The rapid growth of multimodal medical imaging data presents significant storage and transmission challenges, particularly in resource-constrained clinical settings. We propose NEURAL, a novel framework that addresses this by using semantics-guided data compression. Our approach repurposes cross-attention scores between the image and its radiological report from a fine-tuned generative vision-language model to structurally prune chest X-rays, preserving only diagnostically critical regions. This process transforms the image into a highly compressed, graph representation. This unified graph-based representation fuses the pruned visual graph with a knowledge graph derived from the clinical report, creating a universal data structure that simplifies downstream modeling. Validated on the MIMIC-CXR and CheXpert Plus dataset for pneumonia detection, NEURAL achieves a 93.4-97.7\% reduction in image data size while maintaining a high diagnostic performance of 0.88-0.95 AUC, outperforming other baseline models that use uncompressed data. By creating a persistent, task-agnostic data asset, NEURAL resolves the trade-off between data size and clinical utility, enabling efficient workflows and teleradiology without sacrificing performance. Our NEURAL code is available at https://github.com/basiralab/NEURAL.

We present MInDI-3D (Medical Inversion by Direct Iteration in 3D), the first 3D conditional diffusion-based model for real-world sparse-view Cone Beam Computed Tomography (CBCT) artefact removal, aiming to reduce imaging radiation exposure. A key contribution is extending the "InDI" concept from 2D to a full 3D volumetric approach for medical images, implementing an iterative denoising process that refines the CBCT volume directly from sparse-view input. A further contribution is the generation of a large pseudo-CBCT dataset (16,182) from chest CT volumes of the CT-RATE public dataset to robustly train MInDI-3D. We performed a comprehensive evaluation, including quantitative metrics, scalability analysis, generalisation tests, and a clinical assessment by 11 clinicians. Our results show MInDI-3D's effectiveness, achieving a 12.96 (6.10) dB PSNR gain over uncorrected scans with only 50 projections on the CT-RATE pseudo-CBCT (independent real-world) test set and enabling an 8x reduction in imaging radiation exposure. We demonstrate its scalability by showing that performance improves with more training data. Importantly, MInDI-3D matches the performance of a 3D U-Net on real-world scans from 16 cancer patients across distortion and task-based metrics. It also generalises to new CBCT scanner geometries. Clinicians rated our model as sufficient for patient positioning across all anatomical sites and found it preserved lung tumour boundaries well.

Despite the progress of radiology report generation (RRG), existing works face two challenges: 1) The performances in clinical efficacy are unsatisfactory, especially for lesion attributes description; 2) the generated text lacks explainability, making it difficult for radiologists to trust the results. To address the challenges, we focus on a trustworthy RRG model, which not only generates accurate descriptions of abnormalities, but also provides basis of its predictions. To this end, we propose a framework named chain of diagnosis (CoD), which maintains a chain of diagnostic process for clinically accurate and explainable RRG. It first generates question-answer (QA) pairs via diagnostic conversation to extract key findings, then prompts a large language model with QA diagnoses for accurate generation. To enhance explainability, a diagnosis grounding module is designed to match QA diagnoses and generated sentences, where the diagnoses act as a reference. Moreover, a lesion grounding module is designed to locate abnormalities in the image, further improving the working efficiency of radiologists. To facilitate label-efficient training, we propose an omni-supervised learning strategy with clinical consistency to leverage various types of annotations from different datasets. Our efforts lead to 1) an omni-labeled RRG dataset with QA pairs and lesion boxes; 2) a evaluation tool for assessing the accuracy of reports in describing lesion location and severity; 3) extensive experiments to demonstrate the effectiveness of CoD, where it outperforms both specialist and generalist models consistently on two RRG benchmarks and shows promising explainability by accurately grounding generated sentences to QA diagnoses and images.

In this work, we present GazeLT, a human visual attention integration-disintegration approach for long-tailed disease classification. A radiologist's eye gaze has distinct patterns that capture both fine-grained and coarser level disease related information. While interpreting an image, a radiologist's attention varies throughout the duration; it is critical to incorporate this into a deep learning framework to improve automated image interpretation. Another important aspect of visual attention is that apart from looking at major/obvious disease patterns, experts also look at minor/incidental findings (few of these constituting long-tailed classes) during the course of image interpretation. GazeLT harnesses the temporal aspect of the visual search process, via an integration and disintegration mechanism, to improve long-tailed disease classification. We show the efficacy of GazeLT on two publicly available datasets for long-tailed disease classification, namely the NIH-CXR-LT (n=89237) and the MIMIC-CXR-LT (n=111898) datasets. GazeLT outperforms the best long-tailed loss by 4.1% and the visual attention-based baseline by 21.7% in average accuracy metrics for these datasets. Our code is available at https://github.com/lordmoinak1/gazelt.

Segmenting individual instances of mitochondria from imaging datasets can provide rich quantitative information, but is prohibitively time-consuming when done manually, prompting interest in the development of automated algorithms using deep neural networks. Existing solutions for various segmentation tasks are optimized for either: high-resolution three-dimensional imaging, relying on well-defined object boundaries (e.g., whole neuron segmentation in volumetric electron microscopy datasets); or low-resolution two-dimensional imaging, boundary-invariant but poorly suited to large 3D objects (e.g., whole-cell segmentation of light microscopy images). Mitochondria in whole-cell 3D electron microscopy datasets often lie in the middle ground - large, yet with ambiguous borders, challenging current segmentation tools. To address this, we developed skeleton-oriented object segmentation (SKOOTS) - a novel approach that efficiently segments large, densely packed mitochondria. SKOOTS accurately and efficiently segments mitochondria in previously difficult contexts and can also be applied to segment other objects in 3D light microscopy datasets. This approach bridges a critical gap between existing segmentation approaches, improving the utility of automated analysis of three-dimensional biomedical imaging data. We demonstrate the utility of SKOOTS by applying it to segment over 15,000 cochlear hair cell mitochondria across experimental conditions in under 2 hours on a consumer-grade PC, enabling downstream morphological analysis that revealed subtle structural changes following aminoglycoside exposure - differences not detectable using analysis approaches currently used in the field.

Atrial fibrillation, a common cardiac arrhythmia with rapid and irregular atrial electrical activity, requires accurate left atrial segmentation for effective treatment planning. Recently, deep learning methods have gained encouraging success in left atrial segmentation. However, current methodologies critically depend on the assumption of consistently complete centered left atrium as input, which neglects the structural incompleteness and boundary discontinuities arising from random-crop operations during inference. In this paper, we propose BSA-Net, which exploits an adaptive adjustment strategy in both feature position and loss optimization to establish long-range feature relationships and strengthen robust intermediate feature representations in boundary regions. Specifically, we propose a Spatial-adaptive Convolution (SConv) that employs a shuffle operation combined with lightweight convolution to directly establish cross-positional relationships within regions of potential relevance. Moreover, we develop the dual Boundary Prioritized loss, which enhances boundary precision by differentially weighting foreground and background boundaries, thus optimizing complex boundary regions. With the above technologies, the proposed method enjoys a better speed-accuracy trade-off compared to current methods. BSA-Net attains Dice scores of 92.55%, 91.42%, and 84.67% on the LA, Utah, and Waikato datasets, respectively, with a mere 2.16 M parameters-approximately 80% fewer than other contemporary state-of-the-art models. Extensive experimental results on three benchmark datasets have demonstrated that BSA-Net, consistently and significantly outperforms existing state-of-the-art methods.

Arterial input function (AIF) extraction is a crucial step in quantitative pharmacokinetic modeling of DCE-MRI. This work proposes a robust deep learning model that can precisely extract an AIF from DCE-MRI images. A diverse dataset of human brain DCE-MRI images from 289 participants, totaling 384 scans, from five different institutions with extracted gadolinium-based contrast agent curves from large penetrating arteries, and with most data collected for blood-brain barrier (BBB) permeability measurement, was retrospectively analyzed. A 3D UNet model was implemented and trained on manually drawn AIF regions. The testing cohort was compared using proposed AIF quality metric AIFitness and K^trans values from a standard DCE pipeline. This UNet was then applied to a separate dataset of 326 participants with a total of 421 DCE-MRI images with analyzed AIF quality and K^trans values. The resulting 3D UNet model achieved an average AIFitness score of 93.9 compared to 99.7 for manually selected AIFs, and white matter K^trans values were 0.45/min × 10^-3 and 0.45/min × 10^-3, respectively. The intraclass correlation between automated and manual K^trans values was 0.89. The separate replication dataset yielded an AIFitness score of 97.0 and white matter K^trans of 0.44/min × 10^-3. Findings suggest a 3D UNet model with additional convolutional neural network kernels and a modified Huber loss function achieves superior performance for identifying AIF curves from DCE-MRI in a diverse multi-center cohort. AIFitness scores and DCE-MRI-derived metrics, such as K^trans maps, showed no significant differences in gray and white matter between manually drawn and automated AIFs.

Explainable Artificial Intelligence (XAI) encompasses a broad spectrum of methods that aim to enhance the transparency of deep learning models, with Class Activation Mapping (CAM) methods widely used for visual interpretability. However, systematic evaluations of these methods in veterinary radiography remain scarce. This study presents a comparative analysis of eleven CAM methods, including GradCAM, XGradCAM, ScoreCAM, and EigenCAM, on a dataset of 7362 canine and feline X-ray images. A ResNet18 model was chosen based on the specificity of the dataset and preliminary results where it outperformed other models. Quantitative and qualitative evaluations were performed to determine how well each CAM method produced interpretable heatmaps relevant to clinical decision-making. Among the techniques evaluated, EigenGradCAM achieved the highest mean score and standard deviation (SD) of 2.571 (SD = 1.256), closely followed by EigenCAM at 2.519 (SD = 1.228) and GradCAM++ at 2.512 (SD = 1.277), with methods such as FullGrad and XGradCAM achieving worst scores of 2.000 (SD = 1.300) and 1.858 (SD = 1.198) respectively. Despite variations in saliency visualization, no single method universally improved veterinarians' diagnostic confidence. While certain CAM methods provide better visual cues for some pathologies, they generally offered limited explainability and didn't substantially improve veterinarians' diagnostic confidence.

This study aimed to determine if an ensemble (stacking) model that integrates three independently developed base models can reliably predict patients' neurological outcomes following out-of-hospital cardiac arrest (OHCA) within 3 h of arrival and outperform each individual model. This retrospective study included patients with OHCA (≥ 18 years) admitted directly to Nara Medical University between April 2015 and March 2024 who remained comatose for ≥ 3 h after arrival and had suitable head computed tomography (CT) images. The area under the receiver operating characteristic curve (AUC) and Briers scores were used to evaluate the performance of four models (resuscitation-related background OHCA score factors, bilateral pupil diameter, single-slice head CT within 3 h of arrival, and an ensemble stacked model combining these three models) in predicting favourable neurological outcomes at hospital discharge or 1 month, as defined by a Cerebral Performance Category scale of 1-2. Among 533 patients, 82 (15%) had favourable outcomes. The OHCA, pupil, and head CT models yielded AUCs of 0.76, 0.65, and 0.68 with Brier scores of 0.11, 0.13, and 0.12, respectively. The ensemble model outperformed the other models (AUC, 0.82; Brier score, 0.10), thereby supporting its application for early clinical decision-making and optimising resource allocation.

Despite academic success, radiomics-based machine learning algorithms have not reached clinical practice, partially due to limited repeatability/reproducibility. To address this issue, this work aims to identify a stable subset of radiomics features in prostate MRI for radiomics modelling. A prospective study was conducted in 43 patients who received a clinical MRI examination and a research exam with repetition of T2-weighted and two different diffusion-weighted imaging (DWI) sequences with repositioning in between. Radiomics feature (RF) extraction was performed from MRI segmentations accounting for intra-rater and inter-rater effects, and three different image normalization methods were compared. Stability of RFs was assessed using the concordance correlation coefficient (CCC) for different comparisons: rater effects, inter-scan (before and after repositioning) and inter-sequence (between the two diffusion-weighted sequences) variability. In total, only 64 out of 321 (~ 20%) extracted features demonstrated stability, defined as CCC ≥ 0.75 in all settings (5 high-b value, 7 ADC- and 52 T2-derived features). For DWI, primarily intensity-based features proved stable with no shape feature passing the CCC threshold. T2-weighted images possessed the largest number of stable features with multiple shape (7), intensity-based (7) and texture features (28). Z-score normalization for high-b value images and muscle-normalization for T2-weighted images were identified as suitable.