Med Phys [JOURNAL]
Sun
200 papers
Learning ordinal representation across MRI sequences for liver fibrosis staging.
Huo J, Wang Y, Wu N … +3 more , Zhang R, Zhang J, Jin W
Med Phys · 2026 May · PMID 42206568 · Publisher ↗
BACKGROUND: Accurate liver fibrosis staging (LFS) is important for the diagnosis and treatment planning of patients with liver diseases. Noncontrast MRI is suitable for early screening and long-term monitoring in clinica...
BACKGROUND: Accurate liver fibrosis staging (LFS) is important for the diagnosis and treatment planning of patients with liver diseases. Noncontrast MRI is suitable for early screening and long-term monitoring in clinical practice. However, due to the lack of contrast-enhanced details, and given the continuous progression of liver fibrosis and the subtle differences between stages, the use of noncontrast MRI for LFS requires further exploration. PURPOSE: Here, we develop and evaluate a fine-grained deep learning pipeline using noncontrast MRI for precise and improved diagnosis. METHODS: To reflect variations across progression stages, we propose a fine-grained diagnostic model to capture ordinal representation from noncontrast MRI in 450 cases. We employ a multi-scale learning strategy and a learnable attention mechanism to enhance information utilization across noncontrast MRI sequences. Furthermore, we propose a novel hybrid contrastive triplet learning method and a weighted strategy to address the imbalance across progression stages and improve diagnostic performance. RESULTS: The proposed fine-grained diagnostic model achieved an area under the curve (AUC) of 0.877, outperforming the existing LFS model, deep learning baselines, and general MRI-based diagnostic models. For identifying cirrhosis as the endpoint of fibrosis, the AUC further increased to 0.930. The combined use of UMAP clustering and gradient-based attribution methods revealed meaningful feature patterns and predictive mechanisms, demonstrating the potential of the proposed approach for fine-grained diagnosis. CONCLUSIONS: The proposed fine-grained deep learning model effectively leverages multisequence noncontrast MRI for precise LFS and demonstrates improved diagnostic performance.
Efficient vision mamba for MRI super-resolution via hybrid selective scanning.
Safari M, Wang S, Wildman VL … +10 more , Hu M, Eidex Z, Chang CW, Middlebrooks EH, Qiu RLJ, Patel P, Jani AB, Mao H, Tian Z, Yang X
Med Phys · 2026 May · PMID 42204783 · Publisher ↗
BACKGROUND: High-resolution MRI is essential for accurate diagnosis and treatment planning, but its clinical acquisition is often constrained by long scanning times, which increase patient discomfort and reduce scanner t...
BACKGROUND: High-resolution MRI is essential for accurate diagnosis and treatment planning, but its clinical acquisition is often constrained by long scanning times, which increase patient discomfort and reduce scanner throughput. While super-resolution (SR) techniques offer a post-acquisition solution to enhance resolution, existing deep learning approaches face trade-offs between reconstruction fidelity and computational efficiency, limiting their clinical applicability. PURPOSE: This study aims to develop an efficient and accurate deep learning framework for MRI SR that preserves fine anatomical detail while maintaining low computational overhead, enabling practical integration into clinical workflows. MATERIALS AND METHODS: We propose a novel SR framework based on multi-head selective state-space models (MHSSM) integrated with a lightweight channel multilayer perceptron (MLP). The model employs 2D patch extraction with hybrid scanning strategies (vertical, horizontal, and diagonal) to capture long-range dependencies while mitigating pixel forgetting. Each MambaFormer block combines MHSSM, depthwise convolutions, and gated channel mixing to balance local and global feature representation. The framework was trained and evaluated on two distinct datasets: 7T brain T1 MP2RAGE maps (142 subjects) and 1.5T prostate T2w MRI (334 subjects). Performance was compared against multiple baselines including Bicubic interpolation, GAN-based (CycleGAN, Pix2pix, SPSR), transformer-based (SwinIR), Mamba-based (MambaIR), and diffusion-based (ISB, Res-SRDiff) methods. RESULTS: The proposed model demonstrated superior performance across all evaluation metrics while maintaining exceptional computational efficiency. On the 7T brain dataset, our method achieved the highest structural similarity (SSIM: ) and peak signal-to-noise ratio (PSNR: dB), along with the best perceptual quality scores (LPIPS: ; GMSD: ). These results represented statistically significant improvements over all baselines ( ), including a 2.1% SSIM gain over SPSR and a 2.4% PSNR improvement over Res-SRDiff. For the prostate dataset, the model similarly outperformed competing approaches, achieving SSIM of , PSNR of dB, LPIPS of , and GMSD of . Notably, our framework accomplished these results with only 0.9 million parameters and 57 GFLOPs, representing reductions of 99.8% in parameters and 97.5% in computational operations compared to Res-SRDiff, while also substantially outperforming SwinIR and MambaIR in both accuracy and efficiency metrics. CONCLUSION: The proposed framework provides a computationally efficient yet accurate solution for MRI SR, delivering well-defined anatomical details and improved perceptual fidelity across anatomically distinct datasets. By significantly reducing computational demands while maintaining state-of-the-art performance, the model offers strong potential for feasibility toward clinical translation and scalable integration into future imaging workflows.
Impact of LET-modifying planning objectives on the optimization of mixed-modality proton-photon treatments.
Seckler L, Bennan ABA, Wahl N
Med Phys · 2026 May · PMID 42204779 · Publisher ↗
BACKGROUND: The linear energy transfer (LET) correlates with the relative biological effectiveness (RBE); thus, the increase of LET in depth in proton therapy leads to elevated RBE and a higher risk of toxicity in health...
BACKGROUND: The linear energy transfer (LET) correlates with the relative biological effectiveness (RBE); thus, the increase of LET in depth in proton therapy leads to elevated RBE and a higher risk of toxicity in healthy tissue beyond the target. Jointly optimized combined proton-photon treatments may serve as a means to redistribute LET and avoid high-LET regions in organs at risk (OARs). PURPOSE: To mitigate LET-related toxicity in OARs, this work leverages combined treatments by direct integration of LET in the joint optimization process. To this end, LET-modifying objective functions and variable RBE models are used within a joint optimization framework for combined proton-photon treatments. This approach combines the properties of both modalities to achieve LET-shaping and satisfy biological dose objectives in the target and OAR. METHODS: LET-modifying objective (LMO) functions based on the dose-weighted LET (LETxDose) and on dirty dose-defined as high-LET dose above a defined LET-threshold- were integrated into joint optimization within the open source toolkit matRad (v2.10.1). Treatment plans with phantom and clinical test cases were optimized and evaluated, consistently using the same dose objectives in each plan. The compared plans use the following: (1) constant RBE model, (2) variable RBE models, (3) dirty dose objectives, (4) LETxDose (LxD) objectives. Furthermore, sensitivity analysis was performed for various penalty weights, different proton beam angles and - ratios. The plans were assessed for target coverage, dose conformity, and LET reduction in OARs, combining five proton fractions with 25 photon fractions. RESULTS: LET-modifying objectives enabled localized reduction of high-LET exposition in OARs by redistributing dose contributions of both modalities. Protons provided superior tumor-targeted dose delivery and reduced integral dose, while photons mitigated high-LET at the distal edge without compromising dose conformity in the tumor. Beam alignment had a greater influence on enhancing the LMO effect than the number of proton beams. The dirty dose to of brain stem volume was reduced when using joint LET optimization approaches. LET-modifying squared underdosing objectives increased the dirty dose in the GTV by up to compared with the reference joint plan, whereas squared overdosing objectives reduced the dirty dose in the brain stem by . Their combination achieved a increase of dirty dose in the GTV with a corresponding reduction in the brain stem. Despite per fraction, the proton-only plan resulted with a times higher dirty dose per fraction in a joint optimized plan using dirty dose objectives, approximately locally per fraction. Using LETxDose objectives, it was times higher. Dirty dose objectives provided more explicit LET control than LETxDose, with minimal compromise in target coverage. CONCLUSION: Integrating LET-based objectives into a jointly optimized proton-photon system allows for direct steering of the combined LET, improving dose conformity and reducing high-LET exposure in critical regions. This approach leverages modality-specific strengths and offers a practical route toward LET-optimized multimodality radiotherapy for safer, more effective treatments.
Correction to "A proof-of-concept study of multitask learning for cranial synthetic CT generation across heterogeneous MRI field strengths".
Med Phys · 2026 May · PMID 42204773 · Publisher ↗
Abstract loading — click title to view on PubMed.
Vision mamba augmented segment anything model for medical image segmentation.
Li Z, Zhao H, Yin F … +4 more , Zheng L, Xu S, Tie J, Zhao P
Med Phys · 2026 May · PMID 42204760 · Publisher ↗
BACKGROUND: Medical image segmentation is a crucial task for accurate diagnosis and treatment, aiding in the identification of organs and lesions. While SAM has excelled in natural image segmentation, its direct applicat...
BACKGROUND: Medical image segmentation is a crucial task for accurate diagnosis and treatment, aiding in the identification of organs and lesions. While SAM has excelled in natural image segmentation, its direct application to medical images is limited due to significant feature differences. Existing models like MedSAM, despite making progress, face challenges with high computational resource consumption and insufficient accuracy in handling detailed features. PURPOSE: To address the limitations of high computational cost and insufficient segmentation accuracy in existing medical image segmentation models, this study proposes a novel model, VM-MedSAM, designed to be more efficient and precise. METHODS: Inspired by the Mamba architecture, we developed VM-MedSAM. The model incorporates a vision backbone network based on RVM+, freezes the prompt encoder, and optimizes the image encoder from MedSAM. This structural adjustment significantly reduces the number of parameters and improves training efficiency. The proposed model was validated on a medical image dataset covering 12 different abdominal organs. RESULTS: Experimental results demonstrate that VM-MedSAM achieves a slight improvement in abdominal organ segmentation accuracy compared to MedSAM, with significant improvements in lung cancer and brain tumor segmentation. Furthermore, VM-MedSAM reduced the number of parameters by 65.11%, increased training speed by 3.82 times, and decreased model size by 85.41%. CONCLUSIONS: The VM-MedSAM model effectively addresses the challenges of high computational cost and limited accuracy in existing medical image segmentation approaches. Its improved performance and efficiency make it a promising solution for medical image segmentation.
PRISM: An open-source framework for regularized material decomposition on a novel kV dual-layer imager.
de Kermenguy F, Jacobson MW, Sharp GC … +13 more , Myronakis M, Harris TC, Lowther N, Hu YH, Vecchione B, Etemadpour R, Lehmann M, Bruegger R, Birrer V, Arroyo PC, Fueglistaller R, Berbeco RI, Ferguson DM
Med Phys · 2026 May · PMID 42192261 · Publisher ↗
BACKGROUND: A new prototype of kV Dual-Layer Imager enables in-treatment material decomposition. However, conventional decomposition methods, such as direct matrix inversion, result in significant noise amplification, li...
BACKGROUND: A new prototype of kV Dual-Layer Imager enables in-treatment material decomposition. However, conventional decomposition methods, such as direct matrix inversion, result in significant noise amplification, limiting its clinical applicability. PURPOSE: To develop and evaluate a fast, noise-robust material decomposition method for dual-layer kV X-ray imaging. This approach addresses the severe noise amplification and instability associated with direct dual-energy inversion through a regularized, projection-domain real-time framework. METHODS: Dual energy projections were acquired using a prototype kV Dual-Layer Imager on a clinical linear accelerator. A penalized weighted least squares objective function was implemented to perform iterative material decomposition. Four regularization strategies were investigated: quadratic, edge-weighted quadratic, non-local similarity, and the proposed cross-similarity, which enforces consistency in both spatial and spectral domains. Performance was evaluated using a TOR-18FG phantom to quantify the trade-off between noise and spatial resolution (Line Spread Function FWHM) and signal bias, as well as on patient thoracic projections. Computation times were compared between a CPU sparse-matrix implementation and a custom GPU matrix-free implementation. RESULTS: The proposed cross-similarity regularization achieved up to a fivefold noise reduction in water-equivalent images while preserving spatial resolution, outperforming local regularization methods. Unlike standard similarity regularization, which induced signal bias (up to ) with larger search windows, cross-similarity maintained minimal bias (below ). In patient studies, cross-similarity enhanced soft-tissue visibility and preserved fine lung structures better than local methods. The GPU matrix-free implementation achieved decomposition times under , approximately two orders of magnitude faster than the CPU implementation CONCLUSIONS: Our developed method provides a robust framework for high-quality DE material decomposition. The novel cross-similarity regularization offers superior noise suppression and resolution preservation compared to conventional methods. With sub-50 ms processing times, the GPU-accelerated implementation satisfies the latency requirements for real-time clinical applications such as intra-fraction markerless tumor tracking.
Temporal fusion and heatmap regression for precise left ventricular parameter measurement in echocardiographic parasternal long-axis videos.
Chen Y, Shan C, Qi Z … +11 more , Shi Z, Guo G, Wang X, Chen H, Chen F, Fang A, Cheng H, Weng H, Luo S, Yao J, Qian S
Med Phys · 2026 May · PMID 42192259 · Publisher ↗
BACKGROUND: Left ventricular geometric parameters are critical for diagnosing and prognosticating cardiovascular diseases. Currently, most measurement techniques rely on two-dimensional transthoracic echocardiography (TT...
BACKGROUND: Left ventricular geometric parameters are critical for diagnosing and prognosticating cardiovascular diseases. Currently, most measurement techniques rely on two-dimensional transthoracic echocardiography (TTE), where an end-diastolic (ED) frame from the parasternal long-axis (PLAX) view is selected, and key points on the interventricular septum (IVS), left ventricular internal dimension (LVID), and left ventricular posterior wall (LVPW) are identified. However, using a single frame often fails to capture the entire structure of the IVS and LVPW, especially when complex anatomical details or blurred edges are present, leading to positional shifts or loss of key points and, hence, considerable measurement errors. PURPOSE: In this study, we propose an automatic method for measuring left ventricular structural parameters based on echocardiographic PLAX-view videos. The approach focuses on the ED frame along with the immediately preceding and following frames. METHODS: We developed an ultrasound video analysis model that integrates temporally distributed and incomplete structural information to reconstruct the complete anatomies of the IVS and LVPW. The model combines a segmentation branch for precise boundary localization with a heatmap regression branch for chamber centerline and LVID measurement line estimation, enforcing perpendicular anatomical constraints. The dataset comprised 400 PLAX echocardiographic videos from 400 distinct patients, acquired at 56 fps. The data were divided into training and validation sets in a ratio of 8:2. The proposed model was compared with U-Net, U-Net++, DeepLabV3, SegFormer, and TransUNet for segmentation, and HRNet and ViTPose for keypoint detection. Evaluation metrics included mIoU, Dice similarity coefficient (DSC), Hausdorff distance (HD), and average precision ( , , mAP). Statistical significance was assessed using paired t-tests with a significance threshold of , and multiple comparisons were corrected using the Benjamini-Hochberg (BH) procedure. RESULTS: Our results demonstrate robust performance improvements over existing benchmarks. In the segmentation task, our method achieved a mean intersection over union (mIoU) of 83.22% (DSC 0.856, HD 10.174). Statistical analysis demonstrated that this performance is significantly superior to classic models like U-Net ( ), showing a positive small-to-medium effect size ( ). In the keypoint detection task, our approach achieved an mAP of 0.698 ( = 0.965), significantly outperforming the DeepLabV3 baseline ( ) with a positive medium-to-large effect size ( ). Moreover, against strong baselines such as ViTPose, our method maintained a statistically significant advantage ( ) with a positive small effect size ( ). CONCLUSIONS: These outcomes demonstrate the method's robust performance in accurately delineating structural boundaries and reducing measurement errors.
Noisy probing dose facilitated dose prediction for pencil beam scanning proton therapy: Physics enhances generalizability.
Zhang L, Holmes JM, Zhang X … +13 more , Liu Z, Feng H, Li M, Sio TT, Vargas CE, Keole SR, Stützer K, Li S, Liu T, Shen J, Wong WW, Vora SA, Liu W
Med Phys · 2026 May · PMID 42192246 · Full text
BACKGROUND: Accurate and efficient dose calculation is essential for online adaptive planning in proton therapy. Deep learning (DL) has shown promising dose prediction results for pencil beam scanning proton therapy (PBS...
BACKGROUND: Accurate and efficient dose calculation is essential for online adaptive planning in proton therapy. Deep learning (DL) has shown promising dose prediction results for pencil beam scanning proton therapy (PBSPT) in recent years, but existing DL-based dose prediction methods still suffer from limited generalizability and an inability to effectively handle outlier clinical cases. This may lead to inaccurate dose delivery to targets or excessive irradiation to organs at risk (OARs), thereby compromising the safety and efficacy of online adaptive proton therapy. PURPOSE: To design a physics-aware and generalizable AI-based PBSPT dose prediction method that incorporates underlying physics to enhance generalizability, particularly in handling outlier clinical cases. METHODS: This study analyzed PBSPT plans of 103 prostate (93 for training and 10 for testing) and 78 lung cancer patients (68 for training and 10 for testing) from our institution, with each case comprising CT images and structure sets. Using the doses generated by our Monte Carlo-based dose engine as the reference standard, we compared three methods: the region of interest (ROI)-based method, the beam mask and sliding window method, and the proposed noisy probing dose method, which rapidly generates a low-statistics dose via uniformly weighted spots on an expanded spot-placement target volume without optimization. To evaluate the generalizability of these methods to rare treatment planning scenarios, 12 cases with uncommon beam angles or prescription doses were used to assess their performance, which was evaluated using dose-volume histogram (DVH) indices, 3D Gamma passing rates (3%/2 mm/10%), and Dice coefficients for dose agreement, while prediction times were measured to gauge model efficiency. RESULTS: The proposed noisy probing dose method consistently outperformed the ROI-based and beam mask baselines across all evaluated metrics, with more accurate dose agreement and superior generalizability. For DVH indices, the noisy probing dose method achieved the smallest deviation in clinical target volume (CTV) dose coverage: in prostate cancer, CTV D98 deviation was reduced by 45% (from 0.53 ± 0.22 Gy [RBE] for ROI-based) and 29% (from 0.41 ± 0.28 Gy [RBE] for beam mask) to 0.29 ± 0.06 Gy [RBE]; in lung cancer, similar improvements were observed, with CTV D98 deviations reduced to 0.34 ± 0.12 Gy [RBE]. The 3D Gamma passing rates improved to 99.65% ± 1.15% for prostate targets and 97.04% ± 1.17% for lung targets. The dice coefficients of the 90% iso-dose lines were also the highest with the noisy probing dose method (prostate: 0.983 ± 0.005; lung: 0.967 ± 0.01). For the 12 outlier cases, the noisy probing dose method maintained superior generalizability, yielding higher 3D Gamma passing rates (prostate targets: 96.79% ± 0.83%, OARs: 94.29% ± 1.01%; lung targets: 93.38% ± 1.34%, OARs: 93.95% ± 1.32%), demonstrating robust generalizability to rare clinical scenarios. The dose predictions for all testing cases were completed within 0.3 seconds. CONCLUSIONS: A novel noisy probing dose method was proposed for PBSPT dose prediction in prostate and lung cancer patients. By embedding more proton-specific physics, this method demonstrated an improvement in the generalizability of dose prediction.
An experimental method for direct measurement of CT detector presampling MTF from reconstructed images.
Zhan L, Dong F, Chen GH … +1 more , Li K
Med Phys · 2026 May · PMID 42192228 · Publisher ↗
BACKGROUND: Modulation transfer functions (MTFs) measured from reconstructed CT images conventionally represent the combined effects of detector blur, focal spot blur, gantry motion blur, and image reconstruction-induced...
BACKGROUND: Modulation transfer functions (MTFs) measured from reconstructed CT images conventionally represent the combined effects of detector blur, focal spot blur, gantry motion blur, and image reconstruction-induced blur. As a result, direct isolation and measurement of the detector MTF from reconstructed images has long been considered infeasible without explicit knowledge of the individual contributions to the overall CT image MTF. PURPOSE: To develop and validate a new experimental method for measuring the CT detector presampling MTF using reconstructed DICOM images. METHODS: The core idea of the proposed method is to utilize the longitudinal ( ) profile of an edge measured from in the reconstructed image to estimate the edge spread function (ESF), and to use a linear stage to translate the edge along the -axis to achieve oversampling of the ESF. A metallic straight edge was aligned parallel to the axial plane. When static-gantry reconstruction is available, the edge is placed near the surface of the parked detector to minimize focal spot blur; otherwise, it is placed at iso-center and the focal spot blur is measured and corrected. The edge is stepped along in sub-detector-pixel increments, and images from all positions are combined to reconstruct a densely sampled ESF and the detector presampling MTF. The method was tested on three CT systems: a Siemens NAEOTOM Alpha photon-counting detector (PCD) system, a Siemens SOMATOM Force energy-integrating detector (EID) system, and a GE Revolution HD EID system, with validation against manufacturer-reported detector presampling MTFs. RESULTS: The measured detector MTFs showed good agreement with the manufacturer-reported reference values. Both the standard-resolution and ultra-high-resolution (UHR) modes of the PCD exhibited substantially higher spatial resolution than the EID. Specifically, the values for the UHR-mode PCD, standard-mode PCD, Siemens EID, and GE EID were 29.1, 14.1, 9.6, and 9.5 lp/cm, respectively. Importantly, detector MTFs measured using the proposed method were consistent across images acquired with different focal spot sizes and reconstructed using different kernels, demonstrating insensitivity to these confounding system factors. CONCLUSIONS: The proposed method enables accurate measurement of CT detector presampling MTFs directly from reconstructed images by mitigating confounding tomographic reconstruction blur effects, without requiring access to detector counts or projection data.
Development of a practical and high-speed deep learning-based dose calculation model in boron neutron capture therapy for head and neck cancer.
Kato R, Kadoya N, Kato T … +4 more , Takeuchi A, Komori S, Jingu K, Takai Y
Med Phys · 2026 May · PMID 42192222 · Publisher ↗
BACKGROUND: In boron neutron capture therapy (BNCT), Monte Carlo (MC) dose calculations are commonly employed because of the complicated neutron reactions. However, MC dose calculations are generally time-consuming. Rece...
BACKGROUND: In boron neutron capture therapy (BNCT), Monte Carlo (MC) dose calculations are commonly employed because of the complicated neutron reactions. However, MC dose calculations are generally time-consuming. Recently, deep learning (DL)-based dose prediction/calculation has attracted increasing attention; however, the applications of DL models in BNCT are limited and have not been investigated extensively. In addition, there are no practical DL models that can be employed in BNCT clinical practice. PURPOSE: We propose a practical DL model for head and neck cancers using a commercial treatment planning system (TPS) for BNCT. To increase the speed of the MC dose calculations, the proposed DL model converts the BNCT dose components calculated by the coarse dose calculation grid size and low statistical uncertainty in the MC calculation into the dose components calculated under the fine setting. METHODS: In this study, we considered 114 head and neck cancer patients who underwent accelerator-based BNCT at our center. Here, we randomly divided 102 patients for training/validation and 12 patients for testing. The BNCT dose components (i.e., boron, nitrogen, hydrogen, and gamma doses) were calculated for all patients using a commercial TPS for BNCT. We employed the hierarchically dense U-net and converted the BNCT dose components calculated by the coarse setting (grid size/uncertainty = 5 mm/10%) into doses calculated by the fine setting (2 mm/5%). In addition, a physical density map was added to the DL input to improve the conversion accuracy. Taking the fine dose as the ground truth, we evaluated the γ-passing rates with various criteria for each dose component of the coarse and DL doses. The calculation time was also measured in the fine, coarse, and DL doses. RESULTS: In the boron dose, the DL dose exhibited significantly higher γ-passing rates of ≥ 95% with a criterion of 1%/2 mm (dose difference/distance to agreement) than the coarse dose. In the nitrogen and hydrogen doses, the DL dose also demonstrated high γ-passing rates of 95.3% and 94.7% with a criterion of 5%/2 mm. The density map was effective for the hydrogen and nitrogen doses. In addition, the average γ-passing rate with the criterion of 3%/2 mm in the gamma dose achieved 96.2% for the DL dose. The average calculation times for the fine and coarse settings were 984.2 ± 470.2 min and 11.0 ± 2.9 min, respectively, and the average conversion time in the DL model was 0.091 ± 0.020 min. CONCLUSIONS: In this study, the proposed DL model was developed to convert each dose component calculated in the coarse setting to the fine dose to increase the speed of commercial MC dose calculations in BNCT for head and neck cancers. The conversion speed from the coarse dose to the fine dose was considerably rapid, and its performance was highly accurate. The proposed DL model can provide accurate BNCT dose distributions at high speed, thereby contributing to improving the quality of BNCT treatment planning.
Technical Note: Efficacy of CT spatial resolution metrics in clinical settings.
Kawashima H, Salyapongse AM, Toia GV … +2 more , Ichikawa K, Szczykutowicz TP
Med Phys · 2026 May · PMID 42186124 · Publisher ↗
BACKGROUND: The task-based transfer function (TTF) and modulation transfer function (MTF) are spatial resolution metrics in CT; with TTF understood as being more relevant for use on modern nonlinear image reconstruction...
BACKGROUND: The task-based transfer function (TTF) and modulation transfer function (MTF) are spatial resolution metrics in CT; with TTF understood as being more relevant for use on modern nonlinear image reconstruction algorithms. PURPOSE: In clinical practice, images are "magnified" and created at varying reconstruction fields of view. The impact of PACS workstation interpolation and zoomed reconstruction on the TTF and MTF may affect their clinically perceived spatial resolution. This paper studies this potential issue. METHODS: A wire and an iodine rod were imaged to calculate MTF and TTF respectively. MTF values were measured from instances of a wire centered on a pixel and shifted to lay over multiple pixels; analogous to how a clinical detail may either be centered or shifted with respect to the pixel matrix. Images were reconstructed using both low- and high-resolution kernels at reconstruction fields of view (RFOV) 60 to 500 mm. A cadaveric temporal bone was similarly imaged. The effect of PACS interpolation was duplicated by applying bi-cubic interpolation. RESULTS: For the low-resolution kernel, with and without PACS interpolation, the MTF and TTF provided almost consistent spatial resolution measurements only below 300 mm RFOV. For the high-resolution kernel without PACS interpolation, the MTF and TTF measurements were consistent up to approximately 220 mm. With PACS interpolation applied, the centered and shifted MTF diverged, which signaled the images were suffering from pixelation which was confirmed visually in the cadaveric images. CONCLUSION: This paper demonstrates a limitation of using the TTF related to RFOV. The TTF does not account for degradations in image quality easily observed by the human eye (predicted by divergent centered and shifted MTF) due to pixel sampling common to RFOVs used in the clinic.
Simultaneous multi-band multi-spectral imaging using multi-band RF excitation for accelerated metal artifact reduction in MRI-guided interventions.
Luo P, Zhou J, Chen L … +7 more , Wang C, Yuan K, Huang X, Zhou Y, Jiao X, Qi F, Qiu B
Med Phys · 2026 May · PMID 42186114 · Publisher ↗
BACKGROUND: Magnetic resonance imaging (MRI) provides high soft tissue contrast and accurate spatial guidance for interventional procedures. However, metallic needles introduce severe signal voids and phase accumulation...
BACKGROUND: Magnetic resonance imaging (MRI) provides high soft tissue contrast and accurate spatial guidance for interventional procedures. However, metallic needles introduce severe signal voids and phase accumulation that can obscure surrounding anatomy and target lesions, thus compromising procedural accuracy and increasing the risk of injury to adjacent healthy tissue. PURPOSE: To reduce metal-induced artifacts and shorten acquisition time in MRI-guided interventions while improving imaging precision during needle guidance. METHODS: We propose a simultaneous multi-band multi-spectral imaging (SMB-MSI) sequence that uses multi-band radiofrequency pulses, designed from off-resonance field maps estimated using low-resolution two-dimensional multi-spectral imaging (2D MSI) data to acquire multiple spectral bins concurrently. In the current implementation, the proposed SMB-MSI data were acquired from two simultaneously excited bins and reconstructed using direct complex summation.The performance of SMB-MSI was evaluated against fast low-angle shot (FLASH), two-dimensional fast spin echo (2D FSE), and conventional MSI with different numbers of spectral bins using phantom experiments and ex vivo porcine liver models. Puncture needles were imaged at 0.35 T and 1.5 T. Metal-induced artifact width was quantified using the full width at half maximum (FWHM), and image signal-to-noise ratios (SNRs) were assessed in both phantom and ex vivo porcine liver experiments. RESULTS: SMB-MSI markedly reduced artifact width (e.g., 3.8 vs. 5.8 mm and 4.2 vs. 6.6 mm in phantom experiments, and 2.3 vs. 3.7 mm, 2.8 vs. 4.2 mm at 0.35 T and 1.5 T, respectively, relative to five-bin MSI), while shortening acquisition time by approximately 80% compared with five-bin MSI. In addition, SMB-MSI provided higher SNR than conventional MSI in both phantom and ex vivo experiments. Needle tips were clearly visualized with minimal geometric distortion. Consistent improvements were observed in both phantom and ex vivo liver experiments at both field strengths. CONCLUSIONS: SMB-MSI provides rapid and accurate MRI-guided needle visualization with substantially reduced metal-induced artifacts. This approach shows promise for improving the efficiency and precision of MRI-guided interventional procedures.
Pulse-to-pulse analysis of ultra-high dose per pulse electron beams using beam current transformers.
Flores-Mancera MA, Radtke JL, Culberson WS
Med Phys · 2026 May · PMID 42178500 · Publisher ↗
BACKGROUND: Ultra-high dose per pulse (UHDPP) beams have shown to be promising for clinical applications due to their healthy tissue sparing potential. In this radiotherapy modality, a limited number of pulses (often < 2...
BACKGROUND: Ultra-high dose per pulse (UHDPP) beams have shown to be promising for clinical applications due to their healthy tissue sparing potential. In this radiotherapy modality, a limited number of pulses (often < 25) can deliver the prescribed dose, making pulse-to-pulse beam stability essential for accurate dose delivery. In this context, characterizing beam stability prior to metrological or biological studies can reduce the uncertainty in experimental outcomes. PURPOSE: The aim of this work was to provide a framework for analyzing UHDPP beam stability using Faraday-shielded beam current transformers (BCTs) to mitigate downstream charge build-up and ensure accurate real-time monitoring. METHODS: Pulses from 6 and 9 MeV UHDPP electron beams (pulse widths: 0.5-4 µs) produced by an IntraOp® Mobetron® irradiator were measured using two Faraday-shielded BCTs installed in the treatment head. A custom-made scintillator coupled with a silicon-photomultiplier through a fiber optic cable was used as an independent detector. The pulse-to-pulse stability was evaluated across beam configurations varying beam energy, pulse width, pulse repetition frequency, and number of pulses per delivery. Pulse-population normality was assessed using Shapiro-Wilk tests, skewness, and excess kurtosis parameters, and the impact of pulse outliers on delivery uncertainty was quantified. RESULTS: For independent multi-pulse deliveries, pulse populations were generally not normally distributed. Normality was typically observed only after excluding outliers. Distributions were moderately skewed with long tails, consistent with a strong outlier component. Outliers accounted for up to 8% of the pulse population and were more likely to occur among the initial pulses in a delivery. Relative to the mean pulse value, outlier pulses deviated up to 8%, and their contribution to delivery uncertainty increased with mean dose rate. The least unstable beam configurations were identified. Large pulse widths (> 2.0 µs) and pulse repetition frequencies ≤ 30 Hz were associated with reduced beam instability, consistent with reduced outlier influence. This study showed that increasing the number of pulses (> 25) mitigated the relative contribution of outliers, which is particularly relevant in experimental designs using UHDPP beams. CONCLUSIONS: A framework for analyzing UHDPP beam stability was outlined. Normality tests were a useful tool to identify delivery limitations and quantify the influence of outliers. These outliers can significantly affect the prescribed dose in UHDPP conditions, and they should be considered when designing and interpreting future experiments. Several mechanisms that may underlie observed instabilities were also hypothesized.
Construction of a classification model for liver fibrosis in MAFLD based on multiparametric MRI radiomics and machine learning: A rat study.
Xia X, He J, Wen Y … +6 more , Liang Y, Yang M, Pan Y, Li F, Huang Z, Lei P
Med Phys · 2026 May · PMID 42178484 · Publisher ↗
BACKGROUND: Liver fibrosis (LF) severity is an important factor in the clinical management and prognosis of patients with metabolic dysfunction-associated fatty liver disease (MAFLD). Conventional imaging modalities and...
BACKGROUND: Liver fibrosis (LF) severity is an important factor in the clinical management and prognosis of patients with metabolic dysfunction-associated fatty liver disease (MAFLD). Conventional imaging modalities and routine clinical parameters may lack sufficient precision for accurate fibrosis staging. Multiparametric magnetic resonance imaging (mpMRI) provides a noninvasive, quantitative approach that may improve fibrosis assessment. OBJECTIVES: This study aimed to develop and validate a machine learning-based classification model for staging LF severity in MAFLD using mpMRI radiomics in a rat model. MATERIALS AND METHODS: A prospective mpMRI study was conducted on 160 male Sprague-Dawley rats with histologically confirmed LF of varying severity, including healthy controls. Imaging was performed using T2-weighted fat-suppressed (T2-FS) and IDEAL-IQ sequences to derive proton density fat fraction, in-phase, and out-of-phase images. Radiomic features were extracted and screened to identify the most discriminative subset, which were then integrated with six machine learning classifiers-logistic regression (LR), decision tree (DT), support vector machine (SVM), random forest (RF), LightGBM, and AdaBoost, using a 7:3 training-to-validation split. Three models were constructed: conventional, deep learning-based, and hybrid. Diagnostic performance was evaluated using the area under the receiver operating characteristic curve (AUC), decision curve analysis (DCA), net reclassification improvement (NRI), and calibration curves. The optimal model was further assessed with a confusion matrix and benchmarked against the best convolutional neural network (CNN) model. Interpretability was examined via SHapley Additive exPlanations (SHAP) analysis. RESULTS: All three models showed good diagnostic performance for staging LF in metabolic dysfunction-associated steatohepatitis (MASH) and outperformed the CNN model. The hybrid model yielded the highest AUC in distinguishing advanced from non-advanced fibrosis (0.982, 95% CI: 0.947-1.000) with an accuracy of 88%. SHAP analysis indicated that deep learning-derived features made the greatest contribution to predictions and were positively associated with increased risk of MASH and fibrosis progression. CONCLUSION: In a rat model of MAFLD, mpMRI-based radiomics combined with machine learning demonstrated promising diagnostic performance for classifying MASH and staging associated LF. This noninvasive approach may help differentiate advanced from non-advanced in MASH, which may provide supportive information and has potential to support clinical decision-making.
Trading robustness: A scenario-free approach to robust multi-criteria optimization for treatment planning.
Cristoforetti R, Süss P, Becher T … +1 more , Wahl N
Med Phys · 2026 May · PMID 42176318 · Publisher ↗
BACKGROUND: Treatment planning in radiotherapy is inherently a multi-criteria optimization (MCO) problem, as it requires balancing competing clinical goals. Traditionally, the treatment's robustness is not formulated as...
BACKGROUND: Treatment planning in radiotherapy is inherently a multi-criteria optimization (MCO) problem, as it requires balancing competing clinical goals. Traditionally, the treatment's robustness is not formulated as a part of this decision making problem, but dealt with separately through margins or robust optimization. PURPOSE: This work facilitates integration of robustness into multi-criteria optimization using a recently proposed efficient "scenario-free" (s-f) robust optimization approach: Utilizing variance reduction objectives, whose computation is independent of the number of chosen error scenarios, robustness can become part of the multi-criteria decision making process at minimal computational overhead. METHODS: The s-f approach relies on the fast evaluation of the expected dose distribution and mean variance during optimization independent of the scenario number. This is achieved by precomputation of expected dose influence and total variance influence matrices, which can then be used for repeated solving of subproblems in the two explored MCO approaches: Lexicographic Ordering (LO) and full Pareto Front (PF) approximation. Different prioritization strategies within the LO approach are used to assess the impact of variance reduction on the optimization outcome. A 3-objective PF approximation, including a variance reduction objective, is generated to visualize and analyze trade-offs between the competing objectives. The robust optimization is performed including scenarios modeling setup and range errors, as well as organ motion, on 3D- and 4DCT lung cancer patient datasets. Robustness analysis is performed to assess and explore the efficacy of all optimization strategies. RESULTS: The s-f approach enabled robust optimization in MCO with computational times comparable to nominal MCO. Both MCO strategies highlighted the interplay between dosimetric and variance reduction objectives. The LO approach showed how prioritization affects plan quality and robustness, while the PF analysis revealed a clear trade-off between robustness and organ-at-risk sparing. CONCLUSIONS: The proposed s-f robust optimization approach allowed the efficient application of robust MCO by significantly reducing the required computational time. The reported analysis highlighted the conflicting trade-off nature of plan robustness and dosimetric quality, demonstrating how robust MCO supports a more informed and flexible decision-making process in treatment planning.
Improving textural realism in breast phantom images.
Omena LM, de Oliveira GM, do Rêgo TG … +4 more , Barbosa YAM, Teixeira JPV, Filho TMS, Barufaldi B
Med Phys · 2026 May · PMID 42176317 · Publisher ↗
BACKGROUND: Breast phantom images can lack textural realism, limiting their utility for imaging research. Achieving realism across all BI-RADS breast densities is essential for representative, unbiased datasets. PURPOSE:...
BACKGROUND: Breast phantom images can lack textural realism, limiting their utility for imaging research. Achieving realism across all BI-RADS breast densities is essential for representative, unbiased datasets. PURPOSE: This study validated simplex noise for improving realism and density fidelity in simulated mammary parenchyma across BI-RADS categories. METHODS: Two reader trials were conducted with eight students from the Artificial Intelligence Applications Laboratory (ARIA) at the Federal University of Paraíba, Brazil. In the first trial, participants compared paired images with and without simplex noise for overall realism in BI-RADS A (fatty) and BI-RADS D (dense) patterns using 2-AFC. During the second trial, optimized parameters were tested across all four BI-RADS categories. Readers selected the most realistic image in each pair and rated confidence levels. RESULTS: In Trial 1, the simplex enhanced reconstruction was judged as more realistic in 96% of BI-RADS D evaluations, compared to 58.6% for BI-RADS A. Trial 2 achieved a high realism consensus for simplex enhanced reconstructions (1,288 vs. 312 selections), with category improvements: BI-RADS A (92.5%), B (92.25%), C (81%), and D (57.75%). ROC analysis confirmed high discriminative performance, with AUC values of 0.95 (A), 0.74 (B), 0.85 (C), and 0.93 (D). Results indicate density dependent effects, with more pronounced improvements in lower density categories. CONCLUSIONS: The introduction of simplex noise enhances the textural realism of simulated mammographic images, particularly representing variability across BI-RADS categories. These simulations provide a resource for testing and validation without patient data.
Magnetic resonance imaging-guided large-scale lesion patterning with high-intensity focused ultrasound in homogeneous phantoms.
Antoniou A, Evripidou N, Georgiou L … +6 more , Christofi A, Zhao J, Yu L, Li W, Kagadis GC, Damianou C
Med Phys · 2026 May · PMID 42176314 · Publisher ↗
BACKGROUND: In typical high-intensity focused ultrasound (HIFU) therapy, energy is delivered sequentially to multiple focal points to collectively shape the overall treatment zone. However, the behavior of large-scale gr...
BACKGROUND: In typical high-intensity focused ultrasound (HIFU) therapy, energy is delivered sequentially to multiple focal points to collectively shape the overall treatment zone. However, the behavior of large-scale grid sonication patterns in homogeneous media under magnetic resonance imaging (MRI) guidance has not been fully characterized. PURPOSE: To characterize the spatial fidelity and MRI appearance of large-scale HIFU sonication grids in an agar-silica tissue-mimicking phantom and to describe their observed behavior relative to biological tissue. METHODS: Rectangular and irregular large-scale (∼100-point) grids were executed in agar-silica phantoms using an MRI-guided HIFU robotic system with a single-element transducer under 3T MRI guidance. T2-weighted (T2-w) Turbo Spin Echo (TSE) images were acquired after each sonication row for lesion monitoring and post-sonication quantification. Each grid configuration was evaluated in a separate phantom preparation. Descriptive metrics were used to characterize lesion morphology and spatial fidelity, quantified using lesion diameter and Euclidean localization error when individual lesions were distinguishable, and region-level descriptors (area and centroid position) for merged lesion regions. One of the grid protocols was also applied to freshly excised porcine skeletal muscle, with lesion formation monitored using the same MRI-based approach. RESULTS: In the phantom, lesions appeared as hyperintense regions at the intended focal depth and closely reproduced the prescribed grid geometry. For a standardized 10 × 10 grid (4 mm spacing) in the ablative regime, individual lesions exhibited a mean diameter of 3.0 ± 0.7 mm and a mean Euclidean localization error of ∼2.5 ± 0.7 mm relative to the planned coordinates. When inter-point spacing approached the lesion diameter, adjacent sonications produced merged regions that preserved the overall grid footprint while forming contiguous thermal coverage. In the irregular configuration, grid coverage was 96%, with a centroid offset of ∼1 mm. Under the tested conditions, ex vivo porcine muscle exhibited incomplete lesion formation and a shift of ∼15 mm from the prescribed depth, producing an irregular hypointense region displaced toward the tissue surface. CONCLUSIONS: This study documented the successful generation and MRI visualization of large-scale HIFU grid patterns in an agar-silica phantom using T2-w imaging. Under the tested conditions, the observed hyperintense regions showed spatial correspondence with the prescribed grid geometry. Observations in freshly excised porcine skeletal muscle recorded variations in lesion formation in the presence of biological heterogeneity, while the agar-silica phantom provided a standardized platform for system-level testing as configured in this study.
Abdominal body composition assessment with 2D-3D Hybrid Mean-Teacher Network.
Hu P, Zhu Y, Hu J … +5 more , Li X, Yu T, Zhou T, Liang T, Li J
Med Phys · 2026 May · PMID 42176313 · Publisher ↗
BACKGROUND: Human body composition is a key indicator of metabolic health, cardiovascular risk, and other health conditions. Segmentation of different adipose tissues and skeletal muscle based on CT images is a commonly...
BACKGROUND: Human body composition is a key indicator of metabolic health, cardiovascular risk, and other health conditions. Segmentation of different adipose tissues and skeletal muscle based on CT images is a commonly used way to assess body composition accurately. PURPOSE: Most current body composition analysis methods focus on single cross-section CT images, which may have a bias of selecting single-slice images to represent whole body composition. Existing 3D deep-learning-based segmentation methods usually require densely annotated training volumetric data and face challenges of intra- and inter-variability among subjects. This article aims to realize accurate volumetric abdominal wall segmentation and body tissue extraction from CT scans with sparse-labeled training data. METHODS: This study collected 70 CT scans from one public dataset and one private dataset, which were divided into the training (public, ), internal test (public, ) and external test (private, ). About 10% of the slices in training scans are labeled to develop the model and all test scans are delineated slice-by-slice by radiologists. A novel semi-supervised segmentation method based on a 2D-3D hybrid mean-teacher network (2D-3D HMT) and an automated body composition assessment pipeline is proposed. Specifically, a 2D teacher model trained with fully-labeled slices guides the fine-tuning of two 2D and 3D student models, taking advantage of the volumetric model while avoiding intensive slice-by-slice labeling work. In addition, an active contour regularization is incorporated into 3D model to improve boundary smoothness. Then, different adipose and muscle tissues are extracted based on the predicted abdominal wall and other ROIs using specific tissue HU ranges. RESULTS: Experiments were conducted on the internal and external datasets to evaluate the semi-supervised segmentation model and body tissue assessment pipeline. Additionally, the public large-scale AATCT-IDS dataset with adipose tissue annotation was also used to evaluate the proposed method. The proposed method achieved high mean Dice similarity coefficients (DSC) on the internal dataset (0.989, 0.979 and 0.973) and external dataset (0.986, 0.982 and 0.973) for subcutaneous adipose tissue (SAT), visceral adipose tissue (VAT) and skeletal muscle (SM), respectively. The Pearson's correlation analysis was conducted and showed the volumes of SAT, VAT and SM by automated segmentation are strongly correlated with those of manual segmentation ( , , , , , , respectively). Compared to related body component assessment methods, the proposed method reached comparable performance while only using 10% labeled data. CONCLUSIONS: The proposed semi-supervised segmentation method and pipeline can effectively perform body composition assessment with very sparse-labeled CT images.
Monte Carlo modeling of the field size dependency for thin window plane-parallel chambers in kilovoltage X-ray reference dosimetry.
Kadeer F, Healy B, Butler D
Med Phys · 2026 May · PMID 42176312 · Publisher ↗
BACKGROUND: Clinical dosimetry for kV X-ray radiotherapy beams requires calibrated ionization chambers. When the clinical field size differs from the field size used for chamber calibration, in-air measurements result in...
BACKGROUND: Clinical dosimetry for kV X-ray radiotherapy beams requires calibrated ionization chambers. When the clinical field size differs from the field size used for chamber calibration, in-air measurements result in a different magnitude of scatter from the detector stem or body housing affecting the sensitive volume. This results in a field size dependency based on how much of the chamber body is irradiated, which can be corrected for using the factor P. This factor is challenging to characterize, and published data is recommended for this correction. PURPOSE: This study aims to use Monte Carlo (MC) modeled factors as a function of field size and beam quality to establish P correction factor values for the PTW 23342 and 23344 chambers (PTW Freiburg GmbH, Freiburg, Germany) and validate these results with measurements. METHODS: Calculations of the P correction factors were made using the egs_chamber user code of the EGSnrc MC package for the full suite of low energy X-ray beams which Australia's Primary Standards Dosimetry Laboratory (PSDL) use to calibrate these chamber types (20-100 kV, half value layer [HVL] 0.11-6.53 mm Al). The field size diameters ranged from 1 to 12 cm. To validate the modeling, measurements of the field size dependency were conducted on a subset of the Australian PSDL medium energy X-ray beams. The P correction values were normalized to the results corresponding to the 5 cm field size diameter. RESULTS: P correction factors are a function of circular field size diameter and HVL. The MC modeled correction factors agreed with the measured factors, for both chamber types, within the calculated ± 1.7% (k = 2) combined uncertainty. The results in this study agreed within uncertainties with the measurements of Austerlitz et al. (2004) and the IAEA TRS398 CoP (2024) for the PTW 23342 chamber in a 100 kV beam and field size diameters from 3 to 5 cm. However, disagreement up to 3.0% was observed when comparing the results of this study with the IAEA TRS398 CoP (2024) measured data at 10 cm relative to 3 cm, for the 100 kV beam. Higher differences may be attributed to differences in the 100 kV beam qualities, and the unknown scatter contribution effect from the physical chamber holder with the measurement data. It is acknowledged that comparisons with the literature at 100 kV were made at beam qualities which exceeded the recommended upper HVL limit of 2.2 mm Al for these chambers. We consider the MC modeled data to be the more accurate characterization of the P correction factor, since the results are not affected by additional scatter from a chamber holder. Physical measurements would not be immune from this effect and would be variable between holder designs. CONCLUSION: MC modeled P correction factor values for in-air measurements of kilovoltage radiotherapy beams using the PTW 23342 and 23344 ionization chambers were determined and validated with experimental measurements. This work provides field size correction factors for these ionization chambers, improving kV X-ray radiation therapy treatment accuracy with more accurate dosimetry.
Towards quality control and harmonization of deep learning CT radiomics: An in-silico feasibility study with virtual colorectal liver metastases.
Venugopal M, Ramani S, Peoples JJ … +4 more , Do RKG, Simpson AL, Wang G, De Man B
Med Phys · 2026 May · PMID 42169495 · Publisher ↗
BACKGROUND: Radiomic imaging biomarkers are increasingly studied in oncology as a means to support disease prognosis and personalized treatment planning. While deep learning (DL) offers scalable alternatives to handcraft...
BACKGROUND: Radiomic imaging biomarkers are increasingly studied in oncology as a means to support disease prognosis and personalized treatment planning. While deep learning (DL) offers scalable alternatives to handcrafted radiomic features, DL-derived biomarkers are sensitive to variations in image acquisition protocols and scanner hardware-even within a single imaging modality. To ensure reliable and reproducible biomarker estimation, it is essential to (1) provide clinicians with quantitative uncertainty estimates associated with biomarker predictions, and (2) address acquisition-induced variability through harmonization strategies that render consistent performance across diverse imaging conditions. PURPOSE: To evaluate uncertainty estimation as a reliability metric for biomarker prediction and to assess the role of image harmonization in improving cross-scanner inference, we develop deep learning radiomics (DLR) models for joint estimation of biomarkers and associated uncertainties and apply linear harmonization filters to standardize imaging conditions. METHODS: We constructed hybrid digital phantoms by embedding 20,000 virtual colorectal liver metastases into 20 clinical CT liver images to generate metastases-laden simulated scans. The proposed DLR models jointly estimated a selected set of biomarkers and their associated aleatoric uncertainties from simulated images. Models were trained, validated, and tested using a 70:20:10 data split. Variability in CT image acquisition was modelled using two scanner types, two reconstruction kernels, and three x-ray tube current settings. DLR performance was evaluated under direct inference (matched training and test scanners), cross-scanner inference, and harmonized inference using linear harmonization filters. Quantitative evaluation was based on correlation coefficients and root mean squared errors (RMSEs). RESULTS: The estimated uncertainties were consistently higher for biomarker predictions with larger deviations from ground truth. Excluding high-uncertainty predictions improved concordance between predictions and ground truth biomarkers. The proportion of uncertain predictions increased under cross-scanner inference relative to direct inference, indicating reduced biomarker reliability under heterogeneous imaging conditions. Furthermore, application of harmonization filters reduced RMSEs by an average of 43% across biomarkers and experiments during cross-scanner inference, demonstrating improved cross-scanner consistency. CONCLUSIONS: In this controlled in-silico study, uncertainty estimation provided a practical reliability metric for DL-based radiomic biomarker prediction, while image harmonization improved reproducibility across heterogeneous acquisition conditions. These findings demonstrate methodological feasibility within a simulation-based framework but have not yet been validated on clinical CT data, motivating future clinical validation and translation.
About
- Frequency
- Sun
- Papers found
- 200
- RSS feed
- Subscribe