Eye tracking has emerged as a powerful tool for examining visual perception and search strategies in various domains, including medicine. While it is relatively straightforward to apply in 2D settings, its use in 3D medical imaging remains challenging and not yet well explored. This gap is particularly relevant for radiology, where volumetric images such as computed tomography (CT) scans are routinely read by medical experts. Radiologists typically interpret these images by navigating through hundreds of 2D slices, most often viewed in the axial projection. A taxonomy of eye movement data during navigation through a CT volume could be valuable to understand how radiologists approach diagnostic tasks. As an example of the derived taxonomy, we asked two radiologists to search abdominal CTs of the pancreas. We collect eye tracking data and align eye gaze movements with slice navigation to visualize the representation of the pancreas through volume and analyze clinicians' gaze behavior in both space and time.

Collocating deep learning training tasks improves GPU utilization but causes drastic slowdowns due to resource contention and risks Out-of-Memory (OOM) failures. Accurate memory estimation is essential for robust collocation, while GPU utilization – a key proxy for resource contention – enables interference-aware scheduling to reduce slowdowns and improve throughput. Existing GPU memory estimators span three paradigms – analytical models, CPU-side libraries, and ML-based estimators – each with distinct limitations: dependence on detailed model specifications, intrusive integration, poor generalization, and varying latency overhead. GPU heterogeneity further complicates estimation, as identical tasks can exhibit markedly different memory footprints across hardware generations. GPU utilization remains comparatively understudied, further complicated by the non-additive nature of utilization metrics and hardware sensitivity. We conduct a systematic analysis of representative estimators from each paradigm – Horus, PyTorch FakeTensor, and our lightweight ML-based estimator – evaluating accuracy, generalizability, and practical overhead. We construct a synthetic dataset spanning MLPs, CNNs, and Transformers with controlled architectural variations, and train MLP- and Transformer-based estimators for memory prediction. We further experiment with utilization estimation on the same dataset. Our evaluation reveals key tradeoffs and validates estimators against real-world unseen models. Significant challenges remain: analytical models are hardware-dependent, CPU-side libraries impose intrusive integration costs, and ML-based estimators struggle with cross-architecture generalization. We release all datasets, tools, and artifacts to support further research.

Accurate interpretation of volumetric CT requires efficient navigation of 3D image volumes and attention to diagnostically relevant regions. While eye-tracking has been widely studied in 2D medical imaging, its use for expertise assessment in CT settings remains limited. We propose a gaze-informed transformer framework for expertise classification in thoracic CT. Using a DINOv2 backbone, radiologist fixation patterns are integrated into volumetric feature learning through (1) a learnable log-space bias in self-attention and (2) gaze-weighted pooling of patch embeddings. We trained and evaluated our approach on 182 CT reading sessions from five radiologists with varying levels of experience. On a held-out test set, the model achieves an ROC-AUC of 0.91 and F1 score of 0.86, outperforming adapted methods. These findings suggest that incorporating visual search behavior into transformers may support objective, process-based expertise assessment in radiology. Code is available via https://github.com/leiluk1/GazeToSkill.

Pancreatic ductal adenocarcinoma (PDAC) remains one of the most lethal cancers, with survival rates influenced by a variety of factors, including early diagnosis, tumor profile, treatment regimen, and treatment response. The development of PDAC prognostic models is often compromised by incomplete clinical records that need to cover imaging, pathology, surgery, and treatment workflows. We propose an explanation-guided reconstruction framework (xRF) that combines an autoencoder with an ensemble of gradient- and perturbation-based explainability methods to identify and prioritize clinically relevant features during autoencoder training. Such a dual-module architecture ensures that the reconstruction process focuses on features most critical for downstream survival prediction instead of diluting its attention to less relevant but easy-to-reconstruct features. The framework was validated using a cohort of 1531 PDAC patients treated in the Danish Capital Region, with clinical features drawn from CT image readings, surgery and pathology protocols, and chemotherapy records. xRF was validated on four survival prediction tasks: post-diagnosis, post-metastasis, post-surgery, and post-chemotherapy survival. We conducted experiments with synthetic masking levels ranging from 10% to 80% and tested performance on unobserved but clinically meaningful features. The prediction performance of xRF drops by 1-2% for the moderate amount of missing data to 10-15% in the cases with high percentage of features missing when measured against the reference predictions without missing data. These results compare favorably against six alternative machine learning-based reconstruction algorithms.

INTRODUCTION:In early stages of cracked teeth, pulpal inflammation is considered reversible. Cracked teeth with retained pulp vitality demonstrate higher survival rates whereas root canal treatment (RCT) negatively influences survival outcomes. The clinical challenge lies in discerning when RCT is required. This study aimed to develop and validate machine learning models for predicting pulp survival in cracked teeth with reversible pulpitis and investigating associations between patient- and tooth-related variables and treatment outcomes. METHODS:Data from 569 patients with 593 cracked teeth were analyzed using Logistic Regression, Gaussian Processes, Random Forests, and Gradient Boosting. Input features included age, gender, tooth type, and preoperative restorative material. The binary outcome represented pulp survival. A 10-fold stratified nested cross-validation approach was used: the outer loop estimated model performance and the inner loop optimized hyperparameters. Classification thresholds were tuned for probabilistic models. Performance metrics included area under the curve, F1-score, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). RESULTS:Logistic Regression achieved the highest area under the curve (0.64) and F1-score (0.60), with strong PPV (0.76) and moderate NPV (0.48), indicating effective identification of patients requiring RCT. Other models showed similar but slightly lower performance (PPV: 0.74-0.77; NPV: 0.45-0.47). Significant predictors of pulp survival included age and presence of preoperative restorations. Older patients and those with existing restorations were less likely to require RCT. CONCLUSIONS:Machine learning models demonstrated a predictive accuracy of 74%-77%. They could improve diagnostic precision and support timely decision-making for endodontic intervention in the management of cracked teeth with reversible pulpitis.

Objective.Develop and evaluate a patient-level framework for discriminating Hodgkin lymphoma (HL) from non-HL (NHL) on baseline18F-FDG PET by aggregating radiomic descriptors from a variable number of involved lymph nodes in a multi-center setting.Approach.Each patient was modeled as an unordered set of lesion-level PET radiomic feature vectors. PET images were converted to standardized uptake values (SUV), resampled to a uniform voxel spacing of2×2×2mm3, and radiomic features were extracted using PyRadiomics from the original SUV images with a fixed bin width of 0.25 SUV units, following definitions consistent with the Image Biomarker Standardisation Initiative. A total of 107 first-order, texture, and shape features were extracted per lesion and optionally augmented with demographic variables and normalized node center-of-mass coordinates after rigid registration to a common reference space. We compared statistical pooled tabular baselines with learnable set-based aggregation models, including Transformer, Deep Sets, multiple instance learning, Set Transformer, and a graph-based model. Models were trained using five-fold cross-validation on Center 1 (151 patients; 85 HL, 66 NHL) and evaluated on an independent external Center 2 cohort (80 patients; 36 HL, 44 NHL) under direct transfer and unsupervised test-time adaptation (TTA).Main results.Combining radiomics with demographic and spatial features consistently improved discrimination. On Center 1, the best-performing deep aggregation models achieved ROC-AUC (mean±standard deviation) up to0.91±0.05and accuracy up to0.84±0.07. External evaluation on Center 2 revealed a measurable domain shift, with reduced performance under direct transfer. Unsupervised TTA, particularly TENT, improved cross-center generalization for deep aggregation models; for example, the Transformer improved from ROC-AUC0.81±0.05andF1-score0.60±0.23to ROC-AUC0.87±0.01andF1-score0.75±0.01. Feature attribution analysis indicated that both demographic and radiomic descriptors contributed to the predictions.Significance.These results show that multi-nodal PET radiomics combined with demographic and spatial dissemination cues is a promising strategy for patient-level HL versus NHL prediction. The findings also highlight the importance of robustness considerations in multi-center PET radiomics and suggest that TTA may help mitigate cross-center variability.

BackgroundImaging is crucial for the detection of pancreatic diseases. Photon-counting computed tomography (PCCT) is a recent improvement in CT detector technology that may improve pancreatic imaging quality.PurposeTo compare the image quality in pancreatic imaging with PCCT versus conventional CT (EIDCT) both subjectively and quantitatively.Material and MethodsWe retrospectively identified 35 patients scanned with both EIDCT and PCCT in multiple contrast phases. Image quality over 11 peripancreatic areas was rated on a 5-point Likert scale. One reader made quantitative measurements of density and noise. Data analysis was performed using R Studio. Continuous data were compared using a paired t-test and ordinal data with a Wilcoxon signed-rank test.ResultsImage quality was rated significantly higher on PCCT for the pancreatic parenchyma in the late arterial phase (3.87 vs. 2.77; P <0.01), the pancreatic parenchyma in the portal venous phase (3.31 vs. 2.53; P <0.01), pancreatic ducts (2.88 vs. 2.62; P <0.01), superior mesenteric artery (4.10 vs. 2.74; P <0.01), coeliac axis (4.04 vs. 2.70; P <0.01), and portal vein (3.29 vs. 2.52; P <0.01). Noise levels were significantly lower with PCCT with a mean reduction of 5.8 HU across all areas. Dose-length-product was significantly reduced in both the late arterial phase (31.8%; P <0.01) and the portal venous phase (21.5%; P <0.01).ConclusionImage quality was subjectively and quantitatively significantly improved for all evaluated pancreatic and peripancreatic structures with PCCT compared to EIDCT. In addition, radiation dose was significantly reduced.

BackgroundBeam-hardening artefacts and image noise are very common causes of reduced image quality with computed tomography (CT) imaging of the lower pelvis and perirenal area affecting precision of diagnoses in urogenital imaging.PurposeTo investigate whether photon-counting CT (PCCT) improves image quality and/or reduces beam-hardening artefacts in the lower pelvis and in the perirenal area compared to energy-integrating CT (EIDCT).Material and MethodsWe retrospectively identified 35 patients scanned using both EIDCT and PCCT. Four radiologists read both PCCT and EIDCT images. Readers evaluated image quality both subjectively and quantitatively over the left medial perirenal fat and the urinary bladder. Continuous data were compared with a paired t-test and ordinal data with a Wilcoxon signed rank test.ResultsImage quality ratings were higher with PCCT compared to EIDCT. Median scores for the left medial perirenal fat were 5 (interquartile range [IQR] = 4-5) for PCCT and 3 (IQR = 3-4) for EIDCT (P <0.001) and for the lower pelvis 5 (IQR = 4-5) for PCCT and 3 (IQR = 2-4) for EIDCT (P <0.001). Image noise was significantly lower in PCCT scans compared to EIDCT scans (urinary bladder 11.9 HU vs. 17.8 HU, P <0.001; left medial perirenal fat 15.6 HU vs. 22.5 HU, P <0.001). Mean dose-length-product was significantly lower in the PCCT scans with a dose reduction of reduction 21.2% (P <0.001).ConclusionBeam-hardening artefacts were considerable reduced and image noise was significantly lower with PCCT compared to EIDCT at significantly reduced radiation doses. This could have potential implications in the radiological assessment of urogenital diseases.

BACKGROUND AND PURPOSE:To extend the previously reported geometric analysis of HaN-Seg: The Head and Neck Organ-at-Risk CT and MR Segmentation Challenge by integrating a dosimetric evaluation, thereby offering a comprehensive assessment of challenge results with practical insights into their clinical applicability. MATERIALS AND METHODS:Participating teams of the HaN-Seg challenge were tasked to auto-segment 30 organs-at-risk (OARs) in the head and neck region using paired contrast-enhanced computed tomography and T1-weighted magnetic resonance images. The teams were ranked according to their geometric performance, measured by the Dice similarity coefficient (DSC) and 95th-percentile Hausdorff distance (HD95). Here, we extend this evaluation with a forward dosimetric analysis, also known as dosimetric impact approximation, including the verification of OAR dosimetric restriction compliance, assessment of OAR priority ratings, evaluation of segmentation performance relative to tumor proximity, and correlation analysis between geometric and dosimetric metrics. RESULTS:All six teams from the previous geometric analysis were assessed for dosimetric performance on the original 14 test cases. Dosimetric analysis revealed minor performance differences among teams, with the best- and worst-performing teams achieving dosimetric compliance in 70.7% and 67.7% of OAR auto-segmentations, respectively. Most teams successfully met priority 1 dosimetric restrictions including the spinal cord, brainstem, optic chiasm, and optic nerves in 11 out of 14 test cases. The lowest compliance rates were observed for the oral cavity and submandibular glands. Correlation analysis revealed no clear relationship between geometric and dosimetric metrics. CONCLUSION:The high dosimetric compliance highlights the practical utility of deep learning OAR auto-segmentation methods. Lower compliance for the oral cavity and submandibular glands most probably stems from their proximity to tumors and the corresponding steep dose gradients, where certain dosimetric constraints are inherently challenging to meet in clinical practice, or from the limitations of the forward dosimetric analysis. These findings underpin the critical need for both geometric and dosimetric evaluations of OAR auto-segmentation tools to ensure robust validation. Such a comprehensive assessment will be essential as commercial deep learning tools become increasingly integrated into the radiotherapy planning workflow.

While specialized Medical Vision-Language Models (VLMs) have achieved remarkable success in interpreting 2D and 3D medical modalities, their deployment for 3D volumetric data remains constrained by significant computational inefficiencies. Current architectures typically suffer from massive anatomical redundancy due to the direct concatenation of consecutive 2D slices and lack the flexibility to handle heterogeneous information densities across different slices using fixed pruning ratios. To address these challenges, we propose MedPruner, a training-free and model-agnostic hierarchical token pruning framework specifically designed for efficient 3D medical image understanding. MedPruner introduces a two-stage mechanism: an Inter-slice Anchor-based Filtering module to eliminate slice-level temporal redundancy, followed by a Dynamic Information Nucleus Selection strategy that achieves adaptive token-level compression by quantifying cumulative attention weights. Extensive experiments on three 3D medical benchmarks and across three diverse medical VLMs reveal massive token redundancy in existing architectures. Notably, MedPruner enables models such as MedGemma to maintain or even exceed their original performance while retaining fewer than 5

INTRODUCTION:Endoscopic classification of ulcerative colitis (UC) shows high interobserver variation. Previous research demonstrated that artificial intelligence (AI) can match the accuracy of central reading in scoring still images. We now extend this assessment to longer colon segments and integrate AI into clinical workflows, evaluating its use for real-time, video-based classification of disease severity, and as a support system for physicians. METHODS:We trained a convolutional neural network with the Mayo Endoscopic Subscores (MESs) of 2,561 images and 53 videos from 645 patients. The model differentiated scorable from unscorable endoscopy sections through open-set recognition. Validation involved 140 video clips from 44 patients with UC. Six inflammatory bowel disease (IBD) experts and 16 nonexperts rated these videos, with expert scores as the gold standard. We assessed the model's performance and the value as a supporting system. Last, the model underwent an alpha test on a real-world patient as a real-time endoscopic support. RESULTS:The model achieved an accuracy of 82%, with no significant differences between the experts and the AI. When used as a supporting system, it improved non-IBD experts' performance by 12% and disagreed with the primary physician in 20%-39% of cases. During the alpha test, it was successfully integrated into clinical practice, accurately distinguishing between MES 0 and MES 1, consistent with endoscopists' assessments. DISCUSSION:Our innovative AI model shows significant potential for enhancing the accuracy of UC severity classification and improving the proficiency of non-IBD experts. It is designed for clinical use and has proven feasible in real-world testing.

The segmentation of medical images is an important step in various diagnostic applications, including abnormality detection and radiotherapy planning. Recent developments in Artificial Intelligence (AI) have significantly advanced the field of segmentation automation. However, expert-level accuracy has not been achieved for most segmentation tasks, which significantly hampers the adoption of fully-automated medical image segmentation. This paper investigates the idea of efficient correction of medical image segmentation by using not manual controller commands, which can be time-consuming, but gaze movements. We propose a lightweight fine-tuning approach of the Segment Anything Model in medical images, known as MedSAM, to interactively adjust segmentation masks based on gaze point prompts. Our model is specifically trained for the abdominal CT imaging task using the publicly available WORD database. While surpassing state-of-the-art segmentation models, comprehensive studies with medical experts demonstrated that our gaze-assisted interactive approach led to significant improvements in segmentation quality. Specifically, the gaze-assisted corrections increased the average segmentation performance by nearly 62% for difficult medical cases, compared to traditional segmentation methods based on bounding boxes. The main findings of our proposed work include: 1) the substantial improvement in segmentation quality through gaze interaction, 2) the development of an efficient correction mechanism leveraging eye movements, and 3) the demonstration of gaze-assisted segmentation’s superior performance in abdominal imaging tasks. Our innovative approach shows promise for interactive segmentation of medical images and opens the door for further advancements in human-AI interaction in medicine using eye-tracking technology.

Faithful yet compact explanations for vision models remain a challenge, as commonly used dense perturbation masks are often fragmented and overfitted, needing careful post-processing. Here, we present a training-free explanation method that replaces dense masks with smooth tunable contours. A star-convex region is parameterized by a truncated Fourier series and optimized under an extremal preserve/delete objective using the classifier gradients. The approach guarantees a single, simply connected mask, cuts the number of free parameters by orders of magnitude, and yields stable boundary updates without cleanup. Restricting solutions to low-dimensional, smooth contours makes the method robust to adversarial masking artifacts. On ImageNet classifiers, it matches the extremal fidelity of dense masks while producing compact, interpretable regions with improved run-to-run consistency. Explicit area control also enables importance contour maps, yielding a transparent fidelity-area profiles. Finally, we extend the approach to multi-contour and show how it can localize multiple objects within the same framework. Across benchmarks, the method achieves higher relevance mass and lower complexity than gradient and perturbation based baselines, with especially strong gains on self-supervised DINO models where it improves relevance mass by over 15

Landmark detection using chest X-ray images is often a time-consuming task performed by radiologists, followed by the calculation of cardiometric indices and the diagnosis of heart diseases. The goal of this research is to estimate uncertainty in data and knowledge, which leads to increased accuracy and interpretability in models that predict landmark positions and classify human body states. To achieve this, data from about 800 X-ray images labeled by four practicing radiologists was utilized. An uncertaintyaware negative log-likelihood loss was proposed to address both regression and classification problems. The proposed models were trained to predict the mean and log variance of a normal distribution for each input image, with the variance value estimating the uncertainty of the prediction. Natural data uncertainty was also measured from labels provided by different radiologists and implemented in the form of a label transform for the coordinates of the landmarks. The classification models demonstrated good correspondence with methods such as Monte Carlo dropout and deep ensembling, which account for knowledge uncertainty. The proposed uncertainty-aware models achieved the best results in both landmark detection and heart disease diagnosis, with accuracy improvements of 3% and 4.61%, respectively. The proposed methods are effective and easy to use, making their application to other medical problems a subject of interest for further research.

A fundamental barrier to the adoption of AI systems in clinical practice is the insufficient transparency of AI decision-making. The field of Explainable Artificial Intelligence (XAI) seeks to provide human-interpretable explanations for a given AI model. The recently proposed Neural Explanation Mask (NEM) framework is the first XAI method to explain learned representations with high accuracy at real-time speed. NEM transforms a given differentiable model into a self-explaining system by augmenting it with a neural network-based explanation module. This module is trained in an unsupervised manner to output occlusion-based explanations for the original model. However, the current framework does not consider labels associated with the inputs. This makes it unsuitable for many important tasks in the medical domain that require explanations specific to particular output dimensions, such as pathology discovery, disease severity regression, and multi-label data classification. In this work, we address this issue by introducing a loss function for training explanation modules incorporating labels. It steers explanations toward target labels alongside an integrated smoothing operator, which reduces artifacts in the explanation masks. We validate the resulting Neural Explanation Masks with target labels (NEMt) framework on public databases of lung radiographs and skin images. The obtained results are superior to the state-of-the-art XAI methods in terms of explanation relevancy mass, complexity, and sparseness. Moreover, the explanation generation is several hundred times faster, allowing for real-time clinical applications. The code is publicly available at //github.com/baerminator/NEM T.

Medical imaging, particularly radiography, is an indispensable part of diagnosing many chest diseases. Final diagnoses are made by radiologists based on images, but the decision-making process is always associated with a risk of incorrect interpretation. Incorrectly interpreted data can lead to delays in treatment, a prescription of inappropriate therapy, or even a completely missed diagnosis. In this context, our study aims to determine whether it is possible to predict diagnostic errors made by radiologists using eye-tracking technology. For this purpose, we asked 4 radiologists with different levels of experience to analyze 1,000 images covering a wide range of chest diseases. Using eye-tracking data, we calculated the radiologists’ gaze fixation points and generated feature vectors based on this data to describe the radiologists’ gaze behavior during image analysis. Additionally, we emulated the process of revealing the read images following radiologists’ gaze data to create a more comprehensive picture of their analysis. Then we applied a recurrent neural network to predict diagnostic errors. Our results showed a 0.7755 ROC AUC score, demonstrating a significant potential for this approach in enhancing the accuracy of diagnostic error recognition.

INTRODUCTION:This systematic review investigates the potential of artificial intelligence (AI) in improving the accuracy and efficiency of prostate-specific membrane antigen positron emission tomography (PSMA PET) scans for detecting metastatic prostate cancer. EVIDENCE ACQUISITION:A comprehensive literature search was conducted across Medline, Embase, and Web of Science, adhering to PRISMA guidelines. Key search terms included "artificial intelligence," "machine learning," "deep learning," "prostate cancer," and "PSMA PET." The PICO framework guided the selection of studies focusing on AI's application in evaluating PSMA PET scans for staging lymph node and distant metastasis in prostate cancer patients. Inclusion criteria prioritized original English-language articles published up to October 2024, excluding studies using non-PSMA radiotracers, those analyzing only the CT component of PSMA PET-CT, studies focusing solely on intra-prostatic lesions, and non-original research articles. EVIDENCE SYNTHESIS:The review included 22 studies, with a mix of prospective and retrospective designs. AI algorithms employed included machine learning (ML), deep learning (DL), and convolutional neural networks (CNNs). The studies explored various applications of AI, including improving diagnostic accuracy, sensitivity, differentiation from benign lesions, standardization of reporting, and predicting treatment response. Results showed high sensitivity (62% to 97%) and accuracy (AUC up to 98%) in detecting metastatic disease, but also significant variability in positive predictive value (39.2% to 66.8%). CONCLUSIONS:AI demonstrates significant promise in enhancing PSMA PET scan analysis for metastatic prostate cancer, offering improved efficiency and potentially better diagnostic accuracy. However, the variability in performance and the "black box" nature of some algorithms highlight the need for larger prospective studies, improved model interpretability, and the continued involvement of experienced nuclear medicine physicians in interpreting AI-assisted results. AI should be considered a valuable adjunct, not a replacement, for expert clinical judgment.

Predicting radiologists' decisions when reading mammograms is a novel way to reduce the number of false positives and false negatives made at breast cancer screening. In this study, we aimed to enhance the accuracy of predicting radiologists' decisions in mammography by leveraging transfer learning. Our dataset comprised 120 digital mammogram cases, each annotated with radiologists' decisions categorized as true positive (TP), false positive (FP), or false negative (FN). We adopted the ResNet50 convolutional neural network (CNN) for our modeling approach, developing two different models. In the first model, ResNet50 was pretrained on the ImageNet dataset, with the initial layers frozen and the remaining layers fine-tuned to adapt to our mammography data. The second model was initialized with ImageNet weights obtained in the first model and further pretrained using the VinDr-Mammo dataset, an open-access large-scale Vietnamese dataset of full-field digital mammograms (FFDM) consisting of 5,000 four-view exams with breast-level assessments and extensive lesion-level annotations. Our transfer learning method improved the decision prediction accuracy by leveraging features from the VinDr-Mammo models.