PurposeAcetabular dysplasia is a three‐dimensional (3D) disorder characterized by regional femoral head undercoverage and altered joint loading. Percentage femoral head coverage (%FHC) enables regional assessment of acetabular coverage, but its relationship with femoral head computed tomography (CT) attenuation, expressed in Hounsfield units (HU), remains unclear. This study evaluated associations between 3D acetabular coverage and femoral head HU values in symptomatic dysplastic hips.MethodsThis retrospective radiologic study included 40 patients with bilateral acetabular dysplastic morphology and complete bilateral CT imaging scheduled for periacetabular osteotomy between 2023 and 2026, resulting in 80 hips. Total and regional %FHC were calculated in the anterolateral (AL), anteromedial (AM), posteromedial (PM) and posterolateral (PL) regions. Lateral %FHC was defined as the mean of AL and PL coverage. Femoral head HU values were measured in 12 regions of interest on three axial slices across the same four quadrants, and lateral HU was defined as the mean of AL and PL values. Associations were assessed using age‐adjusted linear mixed‐effects models with a random intercept for participant.ResultsIntraobserver and interobserver reliability were good to excellent. Age was independently associated with lower HU values across models. Lateral %FHC was positively associated with lateral HU (β = 10.68 HU per 1% increase; 95% confidence interval [CI], 3.76–17.60; p = 0.002). Regionally, %FHC_AL was associated with anterolateral femoral head Hounsfield unit value (UH_AL; β = 10.59; 95% CI, 2.73–18.44; p = 0.009), and %FHC_PL was associated with posterolateral femoral head Hounsfield unit value (UH_PL; β = 7.36; 95% CI, 1.58–13.13; p = 0.013). Total, AM and PM coverage were not associated with corresponding HU values.ConclusionIn this exploratory imaging study, lower lateral, AL and PL 3D acetabular coverage were associated with lower HU values in corresponding femoral head regions. These findings suggest that regional femoral head HU mapping may provide an imaging correlate of spatial acetabular undercoverage in dysplastic hips, although its clinical, diagnostic and prognostic utility requires further validation.Level of EvidenceLevel III, diagnostic study.

This study aimed to develop and validate an integrated model combining radiomics, deep learning (DL) features from pretreatment 2-18fluoro-2-deoxy-D-glucose positron emission tomography/computed tomography, and clinical variables to predict lymphovascular invasion (LVI) and progression-free survival (PFS) in patients with non-small cell lung cancer (NSCLC). Data from 289 patients with clinically T1-3N0M0 NSCLC from Ningbo Mingzhou Hospital were retrospectively analyzed. Patients were randomly divided into a training cohort (n = 231) and an internal validation cohort (n = 58) in an 8:2 ratio. An external validation cohort (n = 93) from Suzhou Hongci Hospital was included. Radiomics features were extracted from manually segmented tumor volumes using PyRadiomics, whereas DL features were extracted from the largest axial tumor slice and its 2 adjacent slices. Feature selection was performed using Spearman correlation, minimum redundancy maximum relevance algorithm, and Least Absolute Shrinkage and Selection Operator regression. The selected features were used to train and compare multiple machine learning classifiers. Subsequently, 7 models (Clinical, Radiomics, DL, Radiomics-Clinical, DL-Clinical, DL-Radiomics, and DL-Radiomics-Clinical) were evaluated based on the area under the curve (AUC) and accuracy. Patients were stratified into high- and low-risk groups using the optimal Youden index for PFS comparison. The deep learning-radiomics-clinical (DLRC) model demonstrated superior performance, with AUCs of 0.91, 0.81, and 0.84 in the training, internal validation, and external validation cohorts, respectively. In subgroup analyses, the DLRC model showed robust performance for LVI prediction, with AUCs of 0.87 (T1), 0.77 (T2-3), 0.85 (adenocarcinoma), and 0.83 (squamous carcinoma). Kaplan–Meier analysis revealed a significantly shorter PFS in the high-risk group than in the low-risk group (P < .05). The integrated DLRC model enables accurate and noninvasive prediction of LVI status and PFS in NSCLC patients, which can guide personalized treatment strategies and aid in identifying candidates for postoperative adjuvant therapy.

3D image display is essential for next-generation volumetric imaging; however, dense depth multiplexing for 3D image projection remains challenging because diffraction-induced cross-talk rapidly increases as the axial image planes get closer. Here, we introduce a 3D display system comprising a digital encoder and a diffractive decoder, which simultaneously projects different images onto multiple target axial planes with high axial resolution. By leveraging multi-layer diffractive wavefront decoding and deep learning-based end-to-end optimization, the system achieves high-fidelity depth-resolved 3D image projection in a snapshot, enabling axial plane separations on the order of a wavelength. The digital encoder leverages a Fourier encoder network to capture multi-scale spatial and frequency-domain features from input images, integrates axial position encoding, and generates a unified phase representation that simultaneously encodes all images to be axially projected in a single snapshot through a jointly-optimized diffractive decoder. We characterized the impact of diffractive decoder depth, output diffraction efficiency, spatial light modulator resolution, and axial encoding density, revealing trade-offs that govern axial separation and 3D image projection quality. We further demonstrated the capability to display volumetric images containing 28 axial slices, as well as the ability to dynamically reconfigure the axial locations of the image planes, performed on demand. Finally, we experimentally validated a two-plane optical prototype using a single-layer physical decoder, demonstrating close agreement between the measured results and the target images. These results establish the diffractive 3D display system as a compact and scalable framework for depth-resolved snapshot 3D image projection, with potential applications in holographic displays, AR/VR interfaces, and volumetric optical computing.

This study aimed to compare full- (360°) and half-rotation (180°) dental cone-beam computed tomography (CBCT) protocols in terms of effective dose (ED) and quantitative image quality for the maxilla and mandible regions. An Alderson Radiation Therapy phantom was imaged using a Hyperion X9 Pro (Cefla, Imola) CBCT device with a 10 × 6cm field of view. Both anatomical regions were acquired using full- and half-rotation protocols. A total of 67 thermoluminescence dosimeters were positioned in the phantom for ED calculations. Quantitative image-quality assessment was performed using axial slices extracted from each volume, and signal-to-noise ratio (SNRs) and contrast-to-noise ratio (CNRs) were calculated. The EDs were 406.33 µSv (full-rotation) and 208.29 µSv (half-rotation) for the maxilla, and 248.94 µSv (full-rotation) and 73.63 µSv (half-rotation) for the mandible. These doses corresponded to dose reductions of 48.74% in the maxilla and 70.42% in the mandible. SNRs and CNRs decreased by 36.69% and 34.74% in the maxilla, and by 49.95% and 50.28% in the mandible, respectively. Dose reduction was more pronounced in the mandible than in the maxilla, and this was accompanied by a greater loss in SNR and CNR. It is concluded that protocol selection should be guided by diagnostic requirements and adherence to the "As Low as Diagnostically Acceptable being Indication-oriented and Patient-specific (ALADAIP) principle.

To develop a fully automated 2D nnU-Net pipeline for multi-class skeletal muscle segmentation (psoas, paraspinal, and abdominal wall) at the third lumbar (L3) vertebral level, and to quantitatively evaluate its diagnostic performance and reliability compared to manual segmentation. A 2D nnU-Net was trained on 164 axial L3 CT slices from the multi-institutional AMOS22 dataset, spanning diverse abdominal pathologies and multivendor imaging. To assess generalizability under severe anatomical distortion, independent external validation was performed in 50 consecutive patients with advanced liver disease from a single institution (January-December 2025; mean age, 63 ± 15 years; 32 women, 18 men), of whom 88% had moderate-to-severe ascites. Model stability was examined by comparing a five-fold ensemble with the best-performing single-fold model. Intra-observer reliability of the manual reference standard was evaluated in a random subset of 30 cases. Inter-observer agreement was additionally assessed using an independent second reader. Performance metrics included the Dice Similarity Coefficient (DSC), Pearson correlation coefficient (r), and Bland-Altman analysis for cross-sectional areas and mean attenuation. The inference workflow was deployed via a custom Streamlit-based graphical user interface (GUI). In this anatomically complex external validation cohort, the 5-fold ensemble 2D nnU-Net achieved an overall mean DSC of 0.937 ± 0.043 (95% CI, 0.925-0.950), with 80% of cases achieving a mean DSC ≥ 0.90. While the mean DSC was statistically comparable to the best single-fold model (0.937, [95% CI, 0.921-0.952], p = 0.736), the ensemble strategy increased the minimum observed DSC (worst-case performance) from 0.720 to 0.822. Class-specific external validation performance for the 5-fold ensemble was highest for the paraspinal muscles (DSC: 0.960; 95% CI, 0.952-0.967), followed by the psoas muscles (DSC: 0.941; 95% CI, 0.927-0.956), and lowest for the anatomically complex abdominal wall muscles (DSC: 0.911; 95% CI, 0.893-0.929). Comparison between the ensemble model and manual segmentation yielded a Pearson correlation of r = 0.955 (p < 0.001) for total skeletal muscle area, with a mean bias of +7.17 cm2. Intra- and inter-observer agreements for the manual reference standard demonstrated correlation coefficients of r = 0.995 and 0.090 for total areas, respectively. The automated pipeline required 3-5 s per case for inference and quantitative reporting, compared to 3-5 min for manual segmentation. In patients with advanced liver disease and substantial anatomical distortion from ascites, an ensemble-based 2D nnU-Net provides high quantitative agreement with manual L3 skeletal muscle segmentation, while mitigating lower-bound (worst-case) errors relative to single-fold models. Integration with a dedicated GUI enables substantial time savings and supports scalable quantitative body composition measurement.

Ethmoid sinus CBCT imaging as a biometric instrument: dataset creation for deep learning identification.

ADHD prediction from individual-space T1 images using a Vision Transformer with a gross-region grid framework.

While multiple sclerosis (MS) is traditionally regarded as restricted to the central nervous system (CNS), an involvement of the peripheral nervous system (PNS) has been previously detected by histopathology and magnetic resonance neurography (MRN). In the CNS, magnetization transfer contrast (MTC) imaging correlates with areas of demyelination in MS. Here, we aim to characterize and quantify peripheral nerve involvement in patients with relapsing-remitting MS (RRMS) by MTC imaging in correlation with demographic, clinical, and electrophysiologic data. Sixty RRMS patients and 60 age- and sex-matched healthy controls prospectively underwent MTC imaging in a 3.0 Tesla MR scanner. Two axial three-dimensional gradient-echo sequences with and without an off-resonance saturation rapid frequency pulse were conducted at the right mid- to distal thigh. Sciatic nerve regions of interest were manually delineated on ten consecutive axial slices with subsequent evaluation of the magnetization transfer ratio (MTR) of the sciatic nerve. Detailed neurologic and electrophysiologic examinations were conducted in all RRMS patients. Sciatic nerve MTR was lower in RRMS patients (27.2±0.5%) than in controls (29.5±0.4%; p=0.0002) and inversely correlated with the duration of symptoms, the expanded disability status scale, and the distal motor latency of the tibial nerve in RRMS patients as well as with age and the body mass index in controls. Sciatic nerve MTR differentiates between RRMS patients and controls and correlates with important clinical and electrophysiologic data suggesting clinical relevance of an MTR decrease in the PNS. Our results provide further evidence of peripheral nerve involvement in RRMS and point towards peripheral co-demyelination.

Fully Automated Deep Learning-Based Pipeline for Evans Index Measurement from Raw 3D MRI.

This study aimed to evaluate the performance of three pre-trained deep learning models (ResNet50, MobileNetV2, and EfficientNetB0) in the detection of metal artefacts (MAs) in axial cone-beam computed tomography (CBCT) slices. Two researchers calibrated, examined and collected 1000 axial CBCT slices having crown and restoration-related MAs (CRS) (n = 200), orthodontic bracket-related MAs (OB) (n = 200), implant plant-related MAs (IMP) (n = 200), Root canal filling-related MAs (RC) (n = 200) and 200 axial slices without artefacts (N) from a single CBCT unit. The image dataset was split at a 70:20:10 ratio for training, validation and testing. Data augmentation was applied to the training data. Three pre-trained models (ResNet50, MobileNetV2 and EfficientNetB0) were trained, validated and tested for the ability to classify data. The best-performing model was externally validated using axial images from different CBCT units from the same institution. EfficientNetB0 was the best-performing model with 0.96 (95% CI: 0.93-0.98) test accuracy. Its precision, recall and F1-score were 0.98 (95% CI: 0.96-0.99), 0.96 (95% CI: 0.93-0.97) and 0.97(95% CI: 0.95-0.99), respectively. EfficientNetB0 showed a significantly higher Area under curve AUC value (0.98, 95% CI: 0.96-0.99) compared to other models. Its average classification time was 7.1 s. Gradient-weighted class activation mapping attention confirmed the focus of the model on MAs in the images. During external validation, EfficientNetB0 showed an accuracy of 0.93 (95% CI: 0.90-0.96) and an AUC-ROC of 0.95 (95% CI: 0.94-0.96). There was no significant difference in the performance metrics of EfficientNetB0 between test data set and external validation dataset. EfficientNetB0 demonstrated consistent and acceptable performance across all evaluation stages for the detection and classification of MAs in axial CBCT slices. The study describes the performance of pre-trained model in the detection and classification of MAs. The model can be used by the clinician as a decision-support and quality-assessment tool in the CBCT imaging workflow.

AI models for medical imaging often fail under dataset shifts and on underrepresented patient subgroups. Detecting out-of-distribution scans-arising from rare pathologies, atypical anatomy, or acquisition artifacts-is therefore essential for robust deployment. We introduce ViTMARE (Vision Transformer Masked Autoencoder Reconstruction Error), a volumetric anomaly-detection pipeline for 3D brain MRI that leverages Vision Transformer Masked AutoEncoders (ViTMAEs) adapted to volumetric data by treating axial slices as input channels. The model is fine-tuned on normal brain volumes and evaluated using a synthetic-lesion generator that produces anatomically plausible abnormalities. During inference, ViTMARE performs multiple reconstructions (N=100) and aggregates binary anomaly masks via majority voting, followed by morphological closing and opening to suppress spurious noise. On a test set of real images with added synthetic anomalies, ViTMARE achieves a median Dice score of 0.793, a median precision of 0.912, and a median recall of 0.748. We present a reproducible pipeline and demonstrate that combining voting-based fusion with morphological postprocessing yields robust voxel-level anomaly detection.

This study aimed to determine whether the cross-sectional area of subcutaneous adipose tissue (SAT) at the 12th thoracic vertebra (T12) vertebral level on routine preoperative chest CT is associated with post-transplant outcomes. We retrospectively analyzed lung transplant recipients who underwent routine pretransplant chest CT imaging. The SAT area at the T12 level, which serves as a reliable and consistently visible surrogate for whole-body adiposity on routine chest CT, was measured on axial CT slices using standardized tissue attenuation thresholds. Primary outcome was a 1-year composite of all-cause mortality, chronic lung allograft dysfunction and severe infections requiring hospitalization. Secondary outcomes included length of stay, ventilator support, grade III primary graft dysfunction and acute rejection. Associations between T12 SAT and outcomes were evaluated using univariable and multivariable Cox regression models to identify predictors. To assess the incremental predictive value of SAT, we compared nested prediction models that combined SAT with clinical variables. Model performance was assessed using the area under the receiver operating characteristic curve (AUC), net reclassification improvement (NRI), and integrated discrimination improvement (IDI). Of 280 patients, 29 (10.4%) experienced the primary composite outcome, and 75 (26.8%) developed secondary outcomes. Kaplan-Meier analysis showed worse survival in recipients with SAT area above the optimal cutoff of 30.05cm² (log-rank P < 0.01), and SAT area remained an independent predictor of the primary outcome. Adding SAT and SATBMI to the baseline model improved discrimination for the primary outcome, with AUC increasing from 0.65(95%CI:0.43-0.85) to 0.75(95%CI:0.63-0.85), with sensitivity of 76% and specificity of 62% (NRI = 0.83, P < 0.01). For secondary outcomes, adding SATBMI increased the AUC from 0.60(95%CI:0.33-0.86) to 0.85(95%CI:0.71-0.97), with sensitivity of 81% and specificity of 88% (NRI = 0.39, P = 0.01). Preoperative SAT quantification at the T12 level may serve as a useful imaging marker for predicting adverse outcomes in lung transplant recipients, potentially improving risk stratification.

Magnetic resonance imaging (MRI) is a cornerstone of non-invasive diagnosis and response monitoring in neuro-oncology, and predictions of spatial tumor progression conditioned on the patients' anatomy are increasingly important. We present a proof-of-principle of personalized spatial tumor progression on MRI through generative AI, focusing on pediatric Diffuse Midline Glioma (DMG). We employed guided Denoising Diffusion Implicit Models (DDIM) to model anatomical tumor growth in pediatric DMGs on MRI. Multiparametric scans from adult (n = 1,251) and pediatric (n = 144) patients from the BraTS23 challenge were used to train a slice-based framework, conditioned on baseline scans and a target tumor size. Repeated image generations produce probabilistic tumor growth maps highlighting likely regions of progression. The realism of the generated MRIs was evaluated quantitatively and qualitatively through expert assessment. Spatial growth predictions were validated against an independent dataset of longitudinal MRI scans from a multi-institutional pre-radiotherapy DMG dataset (n = 178 paired slices). We generated anatomically coherent, patient-specific T2-FLAIR (fluid-attenuated inversion recovery) MRI axial slices. Quantitative measures and expert evaluations confirmed the high quality of the generated images, which trained radiologists were unable to reliably distinguish from real scans (accuracy 0.53 ± 0.03). While radiomic features analyses showed good agreement (83% non-significant features) between synthetic and real images, a classifier detected subtle pixel-wise differences (accuracy of 0.69). Tumor growth probability maps aligned well with true tumor growth observed in follow-up imaging, obtaining a mean continuous DICE score of 0.79 ± 0.13. We present guided DDIMs as a predictive tool for spatial tumor growth, illustrated for the progression of DMGs, that demonstrates potential for its integration in personalized radiotherapy planning. Our comprehensive image quality analysis highlights the importance of carefully evaluating synthetic data and its integration in research and clinical workflows.

Accurate delineation of teeth and jaw bone in cone-beam computed tomography (CBCT) is essential for digital dental diagnosis, treatment planning, and surgical navigation. However, reliable segmentation remains challenging due to imaging noise, blurred boundaries, and the scarcity of annotated dental datasets. Conventional methods often lose boundary precision under such image variability, whereas large-scale generic models struggle to adapt to CBCT characteristics. To address the need for precise segmentation of CBCT images, we propose localization-adaptive dual-branch accurate segmentation (LADAS), an automatic prompt segmentation framework designed to enhance the delineation of teeth and jaw bone in axial CBCT slices. The framework first localizes tooth and jaw regions and then performs refined segmentation using a dual-branch architecture that balances global structural perception and local detail preservation. Furthermore, instead of retraining the entire model, our method employs lightweight adapter modules featuring dual enhancement in both channel and spatial dimensions to transfer general visual knowledge to CBCT imaging characteristics. We annotated 332 dental CBCT scans, allocating 330 for slice-based training and validation, and reserving 2 full volumes for independent evaluation. Experimental results demonstrate that our method achieves a DSC score of 91.43%, outperforming eight existing segmentation methods, while reducing the average processing time from 203.2min (manual) to 16.5min (AI-assisted). These results demonstrate the potential of our method as an effective assistive tool in dentistry workflows.

ObjectivesTo develop and validate a deep learning-based model capable of generating dopamine transporter (DAT) images from early-phase [18F]-FP-CIT positron emission tomography (PET) imaging.Materials and methodsConditional generative adversarial network was trained using 477 dual-phase [¹⁸F]-FP-CIT PET scans acquired with a conventional PET system. The model generated delayed-phase images from early-phase dynamic scans (30–40 min post-injection), using five adjacent axial slices to predict the central delayed-phase slice. The model was evaluated using an internal validation set from the same scanner and an independent prospective validation set from a digital PET system. Striatal binding ratios (SNBR) and inter-subregional ratios were compared using Pearson’s correlation and receiver operating characteristic-area under the curve (AUC) analyses.ResultsGenerated images showed high similarity to real delayed images. SNBRs for the whole striatum correlated strongly between real and generated images (R = 0.93 and 0.90 for internal and independent sets, respectively). Diagnostic performance was comparable with the highest AUC observed in the posterior putamen (real vs generated: 1.00 vs 0.98, p > 0.2). Visual assessments revealed no uninterpretable images, and interpretability did not differ significantly. Diagnostic accuracy of generated images was comparable to that of real images in the internal validation set for detection of abnormality (p = 0.453) and degenerative parkinsonism (DP) (p = 1.000). In the independent validation set, DP detection remained comparable (p = 0.25), whereas real images demonstrated significantly higher accuracy for abnormality detection (p < 0.001).ConclusionDeep learning-based model generated DAT images achieve satisfactory performance in quantitative and visual assessments across internal and independent validation sets.

Safe colorectal cancer (CRC) resection depends on accurate preoperative understanding of mesenteric vascular anatomy, which is highly variable and poorly represented in existing datasets. We curated dual-phase contrast-enhanced abdominal CT (CECT) scans from 60 adults imaged over one year on a Siemens SOMATOM Force scanner (slice thickness 0.75 mm). Arterial phases (25-30 s) and venous phases (55-60 s) were manually segmented in 3D Slicer by an experienced surgeon and verified by a senior colorectal surgeon; without inter-phase registration to preserve native characteristics. The resource includes 60 CECT studies (50 dual-phase, 10 venous only), each containing ~700-900 axial slices. Deliverables comprise CECT volumes, per-structure 3D label masks for major mesenteric arteries and veins. Example figures of overlays and 3D masks are provided. Intended uses include vessel segmentation, vascular-aware surgical planning, and registration research. Limitations include single-institution acquisition, difficulty delineating small vessels <4 mm, absence of inter-phase registration, and altered vascular anatomy from surgeries.

Improving interproximal surface reconstruction via artificial intelligence-based tooth segmentation with crown-root integrated models: A preliminary study.

Evaluation of the reproducibility and factors affecting perfusion measurement in normal pregnancies with single-slice FAIR Arterial Spin Labeling (ASL).

This study evaluates publicly available deep-learning-based lung segmentation models in transplant-eligible patients to determine their performance across disease severity levels, pathology categories, and lung sides, and to identify limitations impacting their use in preoperative planning in lung transplantation. This retrospective study included 32 patients who underwent chest CT scans at Duke University Health System between 2017 and 2019 (a total of 3645 2D axial slices). Patients with standard axial CT scans were selected based on the presence of two or more lung pathologies of varying severity. Lung segmentation was performed using three previously developed deep learning models: Unet-R231, TotalSegmentator, and MedSAM. Performance was assessed using quantitative metrics (volumetric similarity, Dice similarity coefficient, and Hausdorff distance) and a qualitative measure (four-point clinical acceptability scale). Unet-R231 consistently outperformed TotalSegmentator and MedSAM in general for different severity levels and pathology categories ( ). All models showed significant performance declines from mild to moderate-to-severe cases, particularly in volumetric similarity ( ), without significant differences among lung sides or pathology types. Unet-R231 provided the most accurate automated lung segmentation among evaluated models, with TotalSegmentator being a close second, although their performance declined significantly in moderate-to-severe cases, emphasizing the need for specialized model fine-tuning in severe pathology contexts.

Background/Objectives: Automated brain tumor classification from MRI is particularly challenging when restricted to single post-contrast axial T1-weighted slices without volumetric or clinical context. Methods: We present a four-class (glioma, meningioma, pituitary tumor, no tumor) slice-level classification framework that combines a fine-tuned Swin-Tiny Transformer with inverse-frequency class-weighted learning and a prototype SMT-based symbolic auditing layer for post hoc logical consistency checks. All architectures were trained and evaluated under identical preprocessing, augmentation, optimization, and evaluation protocols. Results: On an internal clinical dataset from Bandırma Onyedi Eylül University Hospital (n = 8040 slices), Swin-Tiny achieved 97.42% slice-level accuracy (macro-F1 97.42%, macro-AUC 0.994), exceeding matched convolutional baselines by approximately eight percentage points. Five-fold stratified cross-validation confirmed stability (mean accuracy 97.40% ± 0.28%). Zero-shot evaluation on the independent BRISC-2025 dataset (n = 6000 slices) yielded 94.82% accuracy and macro-AUC 0.97, indicating maintained performance under acquisition-related distribution shift. Per-class metrics were consistently high across tumor types, with residual errors dominated by glioma–meningioma confusion, reflecting known radiologic overlap on single contrast-enhanced T1 slices. The symbolic auditing layer flagged 1.2–2.9% of predictions as constraint-violating; most such cases were borderline but correctly classified, suggesting sensitivity of heuristic thresholds rather than systematic model failure. Conclusions: These findings support the value of hierarchical shifted-window attention for integrating local texture and broader spatial context in slice-level MRI classification. While patient-wise, multimodal, and prospective validation remain necessary for clinical deployment, this study provides a controlled empirical benchmark and a prototype mechanism for post hoc logical auditing in neuro-oncologic imaging.