Medical professionals, especially those in training, often depend on visual reference materials to support an accurate diagnosis and develop pattern recognition skills. However, existing resources may lack the diversity and accessibility needed for broad and effective clinical learning. This paper presents a system designed to generate realistic medical images from textual descriptions and validate their accuracy through a classification model. A pretrained stable diffusion model was fine-tuned using Low-Rank Adaptation (LoRA) on the PathMNIST dataset, consisting of nine colorectal histopathology tissue types. The generative model was trained multiple times using different training parameter configurations, guided by domain-specific prompts to capture meaningful features. To ensure quality control, a ResNet-18 classification model was trained on the same dataset, achieving 99.76% accuracy in detecting the correct label of a colorectal histopathological medical image. Generated images were then filtered using the trained classifier and an iterative process, where inaccurate outputs were discarded and regenerated until they were correctly classified. The highest performing version of the generative model from experimentation achieved an F1 score of 0.6727, with precision and recall scores of 0.6817 and 0.7111, respectively. Some types of tissue, such as adipose tissue and lymphocytes, reached perfect classification scores, while others proved more challenging due to structural complexity. The self-validating approach created demonstrates a reliable method for synthesizing domain-specific medical images because of high accuracy in both the generation and classification portions of the system, with potential applications in both diagnostic support and clinical education. Future work includes improving prompt-specific accuracy and extending the system to other areas of medical imaging.

The high-resolution three-dimensional (3D) images generated with digital breast tomosynthesis (DBT) in the screening of breast cancer offer new possibilities for early disease diagnosis. Early detection is especially important as the incidence of breast cancer increases. However, DBT also presents challenges in terms of poorer results for dense breasts, increased false positive rates, slightly higher radiation doses, and increased reading times. Deep learning (DL) has been shown to effectively increase the processing efficiency and diagnostic accuracy of DBT images. This article reviews the application and outlook of DL in DBT-based breast cancer screening. First, the fundamentals and challenges of DBT technology are introduced. The applications of DL in DBT are then grouped into three categories: diagnostic classification of breast diseases, lesion segmentation and detection, and medical image generation. Additionally, the current public databases for mammography are summarized in detail. Finally, this paper analyzes the main challenges in the application of DL techniques in DBT, such as the lack of public datasets and model training issues, and proposes possible directions for future research, including large language models, multisource domain transfer, and data augmentation, to encourage innovative applications of DL in medical imaging.

Classical biomedical data science models are trained on a single modality and aimed at one specific task. However, the exponential increase in the size and capabilities of the foundation models inside and outside medicine shows a shift toward task-agnostic models using large-scale, often internet-based, data. Recent research into smaller foundation models trained on specific literature, such as programming textbooks, demonstrated that they can display capabilities similar to or superior to large generalist models, suggesting a potential middle ground between small task-specific and large foundation models. This study attempts to introduce a domain-specific multimodal model, Congress of Neurological Surgeons (CNS)-Contrastive Language-Image Pretraining (CLIP), developed for neurosurgical applications, leveraging data exclusively from Neurosurgery Publications. We constructed a multimodal data set of articles from Neurosurgery Publications through PDF data collection and figure-caption extraction using an artificial intelligence pipeline for quality control. Our final data set included 24 021 figure-caption pairs. We then developed a fine-tuning protocol for the OpenAI CLIP model. The model was evaluated on tasks including neurosurgical information retrieval, computed tomography imaging classification, and zero-shot ImageNet classification. CNS-CLIP demonstrated superior performance in neurosurgical information retrieval with a Top-1 accuracy of 24.56%, compared with 8.61% for the baseline. The average area under receiver operating characteristic across 6 neuroradiology tasks achieved by CNS-CLIP was 0.95, slightly superior to OpenAI's Contrastive Language-Image Pretraining at 0.94 and significantly outperforming a vanilla vision transformer at 0.62. In generalist classification, CNS-CLIP reached a Top-1 accuracy of 47.55%, a decrease from the baseline of 52.37%, demonstrating a catastrophic forgetting phenomenon. This study presents a pioneering effort in building a domain-specific multimodal model using data from a medical society publication. The results indicate that domain-specific models, while less globally versatile, can offer advantages in specialized contexts. This emphasizes the importance of using tailored data and domain-focused development in training foundation models in neurosurgery and general medicine.

This study aims to evaluate a deep learning pipeline for detecting clinically significant prostate cancer (csPCa), defined as Gleason Grade Group (GGG) ≥ 2, using biparametric MRI (bpMRI) and compare its performance with radiological reading. The training dataset included 4381 bpMRI cases (3800 positive and 581 negative) across three continents, with 80% annotated using PI-RADS and 20% with Gleason Scores. The testing set comprised 328 cases from the PROSTATEx dataset, including 34% positive (GGG ≥ 2) and 66% negative cases. A 3D nnU-Net was trained on bpMRI for lesion detection, evaluated using histopathology-based annotations, and assessed with patient- and lesion-level metrics, along with lesion volume, and GGG. The algorithm was compared to non-expert radiologists using multi-parametric MRI (mpMRI). The model achieved an AUC of 0.83 (95% CI: 0.80, 0.87). Lesion-level sensitivity was 0.85 (95% CI: 0.82, 0.94) at 0.5 False Positives per volume (FP/volume) and 0.88 (95% CI: 0.79, 0.92) at 1 FP/volume. Average Precision was 0.55 (95% CI: 0.46, 0.64). The model showed over 0.90 sensitivity for lesions larger than 650 mm³ and exceeded 0.85 across GGGs. It had higher true positive rates (TPRs) than radiologists equivalent FP rates, achieving TPRs of 0.93 and 0.79 compared to radiologists' 0.87 and 0.68 for PI-RADS ≥ 3 and PI-RADS ≥ 4 lesions (p ≤ 0.05). The DL model showed strong performance in detecting csPCa on an independent test cohort, surpassing radiological interpretation and demonstrating AI's potential to improve diagnostic accuracy for non-expert radiologists. However, detecting small lesions remains challenging. Question Current prostate cancer detection methods often do not involve non-expert radiologists, highlighting the need for more accurate deep learning approaches using biparametric MRI. Findings Our model outperforms radiologists significantly, showing consistent performance across Gleason Grade Groups and for medium to large lesions. Clinical relevance This AI model improves prostate detection accuracy in prostate imaging, serves as a benchmark with reference performance on a public dataset, and offers public PI-RADS annotations, enhancing transparency and facilitating further research and development.

Bi-parametric magnetic resonance imaging (bpMRI) has become a pivotal modality in the detection and diagnosis of clinically significant prostate cancer (csPCa). Developing AI-based systems to identify csPCa using bpMRI can transform prostate cancer (PCa) management by improving efficiency and cost-effectiveness. However, current state-of-the-art methods using convolutional neural networks (CNNs) and Transformers are limited in learning in-plane and three-dimensional spatial information from anisotropic bpMRI. Their performances also depend on the availability of large, diverse, and well-annotated bpMRI datasets. To address these challenges, we propose the Zonal-aware Self-supervised Mesh Network (Z-SSMNet), which adaptively integrates multi-dimensional (2D/2.5D/3D) convolutions to learn dense intra-slice information and sparse inter-slice information of the anisotropic bpMRI in a balanced manner. We also propose a self-supervised learning (SSL) technique that effectively captures both intra-slice and inter-slice semantic information using large-scale unlabeled data. Furthermore, we constrain the network to focus on the zonal anatomical regions to improve the detection and diagnosis capability of csPCa. We conducted extensive experiments on the PI-CAI (Prostate Imaging - Cancer AI) dataset comprising 10000+ multi-center and multi-scanner data. Our Z-SSMNet excelled in both lesion-level detection (AP score of 0.633) and patient-level diagnosis (AUROC score of 0.881), securing the top position in the Open Development Phase of the PI-CAI challenge and maintained strong performance, achieving an AP score of 0.690 and an AUROC score of 0.909, and securing the second-place ranking in the Closed Testing Phase. These findings underscore the potential of AI-driven systems for csPCa diagnosis and management.

Accurate segmentation of hepatic vessels is pivotal for guiding preoperative planning in ablation surgery utilizing CT images. While non-contrast CT images often lack observable vessels, we focus on segmenting hepatic vessels within preoperative MR images. However, the vascular structures depicted in MR images are susceptible to noise, leading to challenges in connectivity. To address this issue, we propose a two-stage two-stream graph attention U-Net (i.e., TTGA U-Net) for hepatic vessel segmentation. Specifically, the first-stage network employs a CNN or Transformer-based architecture to preliminarily locate the vessel position, followed by an improved superpixel segmentation method to generate graph structures based on the positioning results. The second-stage network extracts graph node features through two parallel branches of a graph spatial attention network (GAT) and a graph channel attention network (GCT), employing self-attention mechanisms to balance these features. The graph pooling operation is utilized to aggregate node information. Moreover, we introduce a feature fusion module instead of skip connections to merge the two graph attention features, providing additional information to the decoder effectively. We establish a novel well-annotated high-quality MR image dataset for hepatic vessel segmentation and validate the vessel connectivity enhancement network's effectiveness on this dataset and the public dataset 3D IRCADB. Experimental results demonstrate that our TTGA U-Net outperforms state-of-the-art methods, notably enhancing vessel connectivity.

Autism spectrum disorder (ASD) is a neurodevelopmental condition that affects a child's cognitive and social skills, often diagnosed only after symptoms appear around age 2. Leveraging MRI for early ASD detection can improve intervention outcomes. This study proposes a framework for autism detection and region localization using an optimized deep learning approach with attention mechanisms. The pipeline includes MRI image collection, pre-processing (bias field correction, histogram equalization, artifact removal, and non-local mean filtering), and autism classification with a Symmetric Structured MobileNet with Attention Mechanism (SSM-AM). Enhanced by Refreshing Awareness-aided Election-Based Optimization (RA-EBO), SSM-AM achieves robust classification. Abnormality region localization utilizes a Multiscale Dilated Attention-based Adaptive U-Net (MDA-AUnet) further optimized by RA-EBO. Experimental results demonstrate that our proposed model outperforms existing methods, achieving an accuracy of 97.29%, sensitivity of 97.27%, specificity of 97.36%, and precision of 98.98%, significantly improving classification and localization performance. These results highlight the potential of our approach for early ASD diagnosis and targeted interventions. The datasets utilized for this work are publicly available at https://fcon_1000.projects.nitrc.org/indi/abide/.

Atrial fibrillation (AF) is the most common cardiac arrhythmia, affecting up to 2 % of the population. Catheter ablation is a promising treatment for AF, particularly for paroxysmal AF patients, but it often has high recurrence rates. Developing in silico models of patients' atria during the ablation procedure using cardiac MRI data may help reduce these rates. This study aims to develop an effective automated deep learning-based segmentation pipeline by compiling a specialized dataset and employing standardized labeling protocols to improve segmentation accuracy and efficiency. In doing so, we aim to achieve the highest possible accuracy and generalization ability while minimizing the burden on clinicians involved in manual data segmentation. We collected LGE-MRI data from VMRC and the cDEMRIS database. Two specialists manually labeled the data using standardized protocols to reduce subjective errors. Neural network (nnU-Net and smpU-Net++) performance was evaluated using statistical tests, including sensitivity and specificity analysis. A new database of LGE-MRI images, based on manual segmentation, was created (VMRC). Our approach with consistent labeling protocols achieved a Dice coefficient of 92.4 % ± 0.8 % for the cavity and 64.5 % ± 1.9 % for LA walls. Using the pre-trained RIFE model, we attained a Dice score of approximately 89.1 % ± 1.6 % for atrial LGE-MRI imputation, outperforming classical methods. Sensitivity and specificity values demonstrated substantial enhancement in the performance of neural networks trained with the new protocol. Standardized labeling and RIFE applications significantly improved machine learning tool efficiency for constructing 3D LA models. This novel approach supports integrating state-of-the-art machine learning methods into broader in silico pipelines for predicting ablation outcomes in AF patients.

Kidney irregularities pose a significant public health challenge, often leading to severe complications, yet the limited availability of nephrologists makes early detection costly and time-consuming. To address this issue, we propose a deep learning framework for automated kidney disease detection, leveraging feature fusion and sequential modeling techniques to enhance diagnostic accuracy. Our study thoroughly evaluates six pretrained models under identical experimental conditions, identifying ResNet50 and VGG19 as the highly efficient models for feature extraction due to their deep residual learning and hierarchical representations. Our proposed methodology integrates feature fusion with an inception block to extract diverse feature representations while maintaining imbalance dataset overhead. To enhance sequential learning and capture long-term dependencies in disease progression, ConvLSTM is incorporated after feature fusion. Additionally, Inception block is employed after ConvLSTM to refine hierarchical feature extraction, further strengthening the proposed model ability to leverage both spatial and temporal patterns. To validate our approach, we introduce a new named Multiple Hospital Collected (MHC-CT) dataset, consisting of 1860 tumor and 1024 normal kidney CT scans, meticulously annotated by medical experts. Our model achieves 99.60 % accuracy on this dataset, demonstrating its robustness in binary classification. Furthermore, to assess its generalization capability, we evaluate the model on a publicly available benchmark multiclass CT scan dataset, achieving 91.31 % accuracy. The superior performance is attributed to the effective feature fusion using inception blocks and the sequential learning capabilities of ConvLSTM, which together enhance spatial and temporal feature representations. These results highlight the efficacy of the proposed framework in automating kidney disease detection, providing a reliable, and efficient solution for clinical decision-making. https://github.com/VS-EYE/KidneyDiseaseDetection.git.

A cardiac digital twin is a virtual replica of a patient-specific heart, mimicking its anatomy and physiology. A crucial step of building a cardiac digital twin is anatomical twinning, where the computational mesh of the digital twin is tailored to the patient-specific cardiac anatomy. In a number of studies, the effect of anatomical variation on clinically relevant functional measurements like electrocardiograms (ECGs) is investigated, using computational simulations. While such a simulation environment provides researchers with a carefully controlled ground truth, the impact of anatomical differences on functional measurements in real-world patients remains understudied. In this study, we develop a biventricular statistical shape model and use it to quantify the effect of biventricular anatomy on ECG-derived and demographic features, providing novel insights for the development of digital twins of cardiac electrophysiology. To this end, a dataset comprising high-resolution cardiac CT scans from 271 healthy individuals, including athletes, is utilized. Furthermore, a novel, universal, ventricular coordinate-based method is developed to establish lightweight shape correspondence. The performance of the shape model is rigorously established, focusing on its dimensionality reduction capabilities and the training data requirements. The most important variability in healthy ventricles captured by the model is their size, followed by their elongation. These anatomical factors are found to significantly correlate with ECG-derived and demographic features. Additionally, a comprehensive synthetic cohort is made available, featuring ready-to-use biventricular meshes with fiber structures and anatomical region annotations. These meshes are well-suited for electrophysiological simulations.