The U-shaped encoder-decoder architecture with skip connections has become a prevailing paradigm in medical image segmentation due to its simplicity and effectiveness. While many recent works aim to improve this framework by designing more powerful encoders and decoders, employing advanced convolutional neural networks (CNNs) for local feature extraction, Transformers or state space models (SSMs) such as Mamba for global context modeling, or hybrid combinations of both, these methods often struggle to fully utilize pretrained vision backbones (e.g., ResNet, ViT, VMamba) due to structural mismatches. To bridge this gap, we introduce ACM-UNet, a general-purpose segmentation framework that retains a simple UNet-like design while effectively incorporating pretrained CNNs and Mamba models through a lightweight adapter mechanism. This adapter resolves architectural incompatibilities and enables the model to harness the complementary strengths of CNNs and SSMs-namely, fine-grained local detail extraction and long-range dependency modeling. Additionally, we propose a hierarchical multi-scale wavelet transform module in the decoder to enhance feature fusion and reconstruction fidelity. Extensive experiments on the Synapse and ACDC benchmarks demonstrate that ACM-UNet achieves state-of-the-art performance while remaining computationally efficient. Notably, it reaches 85.12% Dice Score and 13.89mm HD95 on the Synapse dataset with 17.93G FLOPs, showcasing its effectiveness and scalability. Code is available at: https://github.com/zyklcode/ACM-UNet.

Recent advances in deep learning-based medical image registration have shown that training deep neural networks~(DNNs) does not necessarily require medical images. Previous work showed that DNNs trained on randomly generated images with carefully designed noise and contrast properties can still generalize well to unseen medical data. Building on this insight, we propose using registration between random images as a proxy task for pretraining a foundation model for image registration. Empirical results show that our pretraining strategy improves registration accuracy, reduces the amount of domain-specific data needed to achieve competitive performance, and accelerates convergence during downstream training, thereby enhancing computational efficiency.

To assess the prognostic utility of coronary artery calcium (CAC) scoring and coronary computed tomography angiography (CCTA)-derived quantitative plaque metrics for predicting adverse cardiovascular outcomes. The study enrolled 2404 patients with suspected coronary artery disease (CAD) but without a prior history of CAD. All participants underwent CAC scoring and CCTA, with plaque metrics quantified using an artificial intelligence (AI)-based tool (Cleerly, Inc). Percent atheroma volume (PAV) and non-calcified plaque volume percentage (NCPV%), reflecting total plaque burden and the proportion of non-calcified plaque volume normalized to vessel volume, were evaluated. The primary endpoint was a composite of all-cause mortality and non-fatal myocardial infarction (MI). Cox proportional hazard models, adjusted for clinical risk factors and early revascularization, were employed for analysis. During a median follow-up of 7.0 years, 208 patients (8.7%) experienced the primary endpoint, including 73 cases of MI (3%). The model incorporating PAV demonstrated superior discriminatory power for the composite endpoint (AUC = 0.729) compared to CAC scoring (AUC = 0.706, P = 0.016). In MI prediction, PAV (AUC = 0.791) significantly outperformed CAC (AUC = 0.699, P < 0.001), with NCPV% showing the highest prognostic accuracy (AUC = 0.814, P < 0.001). AI-driven assessment of coronary plaque burden enhances prognostic accuracy for future adverse cardiovascular events, highlighting the critical role of comprehensive plaque characterization in refining risk stratification strategies.

Gastrointestinal (GI) endoscopy is essential in identifying GI tract abnormalities in order to detect diseases in their early stages and improve patient outcomes. Although deep learning has shown success in supporting GI diagnostics and decision-making, these models require curated datasets with labels that are expensive to acquire. Foundation models offer a promising solution by learning general-purpose representations, which can be finetuned for specific tasks, overcoming data scarcity. Developing foundation models for medical imaging holds significant potential, but the sensitive and protected nature of medical data presents unique challenges. Foundation model training typically requires extensive datasets, and while hospitals generate large volumes of data, privacy restrictions prevent direct data sharing, making foundation model training infeasible in most scenarios. In this work, we propose a FL framework for training foundation models for gastroendoscopy imaging, enabling data to remain within local hospital environments while contributing to a shared model. We explore several established FL algorithms, assessing their suitability for training foundation models without relying on task-specific labels, conducting experiments in both homogeneous and heterogeneous settings. We evaluate the trained foundation model on three critical downstream tasks--classification, detection, and segmentation--and demonstrate that it achieves improved performance across all tasks, highlighting the effectiveness of our approach in a federated, privacy-preserving setting.

Accurate vertebrae segmentation is crucial for modern surgical technologies, and deep learning networks provide valuable tools for this task. This study explores the application of advanced deep learning-based methods for segmenting vertebrae in computed tomography (CT) images of adolescent idiopathic scoliosis (AIS) patients. In this study, we collected a dataset of 31 samples from AIS patients, covering a wide range of spinal regions from cervical to lumbar vertebrae. High-resolution CT images were obtained for each sample, forming the basis of our segmentation analysis. We utilized 2 popular neural networks, U-Net and Attention U-Net, to segment the vertebrae in these CT images. Segmentation performance was rigorously evaluated using 2 key metrics: the Dice Coefficient Score to measure overlap between segmented and ground truth regions, and the Hausdorff distance (HD) to assess boundary dissimilarity. Both networks performed well, with U-Net achieving an average Dice coefficient of 92.2 ± 2.4% and an HD of 9.80 ± 1.34 mm. Attention U-Net showed similar results, with a Dice coefficient of 92.3 ± 2.9% and an HD of 8.67 ± 3.38 mm. When applied to the challenging anatomy of AIS, our findings align with literature results from advanced 3D U-Nets on healthy spines. Although no significant overall difference was observed between the 2 networks (P > .05), Attention U-Net exhibited an improved Dice coefficient (91.5 ± 0.0% vs 88.8 ± 0.1%, P = .151) and a significantly better HD (9.04 ± 4.51 vs. 13.60 ± 2.26 mm, P = .027) in critical scoliosis sites (mid-thoracic region), suggesting enhanced suitability for complex anatomy. Our study indicates that U-Net neural networks are feasible and effective for automated vertebrae segmentation in AIS patients using clinical 3D CT images. Attention U-Net demonstrated improved performance in thoracic levels, which are primary sites of scoliosis and may be more suitable for challenging anatomical regions.

Imaging is crucial to assess disease extent, activity, and outcomes in inflammatory bowel disease (IBD). Artificial intelligence (AI) image interpretation requires automated exploitation of studies at scale as an initial step. Here we evaluate natural language processing to classify Crohn's disease (CD) on CTE. From our population representative IBD registry a sample of CD patients (male: 44.6%, median age: 50 IQR37-60) and controls (n = 981 each) CTE reports were extracted and split into training- (n = 1568), development- (n = 196), and testing (n = 198) datasets each with around 200 words and balanced numbers of labels, respectively. Predictive classification was evaluated with CNN, Bi-LSTM, BERT-110M, LLaMA-3.3-70B-Instruct and DeepSeek-R1-Distill-LLaMA-70B. While our custom IBDBERT finetuned on expert IBD knowledge (i.e. ACG, AGA, ECCO guidelines), outperformed rule- and rationale extraction-based classifiers (accuracy 88.6% with pre-tuning learning rate 0.00001, AUC 0.945) in predictive performance, LLaMA, but not DeepSeek achieved overall superior results (accuracy 91.2% vs. 88.9%, F1 0.907 vs. 0.874).

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. 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. 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%). 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.

Basal ganglia intracranial hemorrhage (bgICH) morphology is associated with postoperative functional outcomes. We hypothesized that bgICH spatial representation modeling could be automated for functional outcome prediction after minimally invasive surgical (MIS) evacuation. A training set of 678 computed tomography head and computed tomography angiography images from 63 patients were used to train key-point detection and instance segmentation convolutional neural network-based models for anatomic landmark identification and bgICH segmentation. Anatomic landmarks included the bilateral orbital rims at the globe's maximum diameter and the posterior-most aspect of the tentorial incisura, which were used to define a universal stereotactic reference frame across patients. Convolutional neural network models were tested using volumetric computed tomography head/computed tomography angiography scans from 45 patients who underwent MIS bgICH evacuation with recorded modified Rankin Scales within one year after surgery. bgICH volumes were highly correlated (R2 = 0.95, P < .001) between manual (median 39-mL) and automatic (median 38-mL) segmentation methods. The absolute median difference between groups was 2-mL (IQR: 1-6 mL). Median localization accuracy (distance between automated and manually designated coordinate frames) was 4 mm (IQR: 3-6). Landmark coordinates were highly correlated in the x- (medial-lateral), y- (anterior-posterior), and z-axes (rostral-caudal) for all 3 landmarks (R2 range = 0.95-0.99, P < .001 for all). Functional outcome (modified Rankin Scale 4-6) was predicted with similar model performance using automated (area under the receiver operating characteristic curve = 0.81, 95% CI: 0.67-0.94) and manually (area under the receiver operating characteristic curve = 0.84, 95% CI: 0.72-0.96) constructed spatial representation models (P = .173). Computer vision models can accurately replicate bgICH manual segmentation, stereotactic localization, and prognosticate functional outcomes after MIS bgICH evacuation.

We present results of the development of Dentomo artificial intelligence model based on two neural networks. The model includes a database and a knowledge base harmonized with SNOMED CT that allows processing and interpreting the results of cone beam computed tomography (CBCT) scans of the dental system, in particular, identifying and classifying teeth, identifying CT signs of pathology and previous treatments. Based on these data, artificial intelligence can draw conclusions and generate medical reports, systematize the data, and learn from the results. The diagnostic effectiveness of Dentomo was evaluated. The first results of the study have demonstrated that the model based on neural networks and artificial intelligence is a valuable tool for analyzing CBCT scans in clinical practice and optimizing the dentist workflow.

Deep learning has become the state-of-the-art approach to medical tomographic imaging. A common approach is to feed the result of a simple inversion, for example the backprojection, to a multiscale convolutional neural network (CNN) which computes the final reconstruction. Despite good results on in-distribution test data, this often results in overfitting certain large-scale structures and poor generalization on out-of-distribution (OOD) samples. Moreover, the memory and computational complexity of multiscale CNNs scale unfavorably with image resolution, making them impractical for application at realistic clinical resolutions. In this paper, we introduce GLIMPSE, a local coordinate-based neural network for computed tomography which reconstructs a pixel value by processing only the measurements associated with the neighborhood of the pixel. GLIMPSE significantly outperforms successful CNNs on OOD samples, while achieving comparable or better performance on in-distribution test data and maintaining a memory footprint almost independent of image resolution; 5GB memory suffices to train on 1024 × 1024 images which is orders of magnitude less than CNNs. GLIMPSE is fully differentiable and can be used plug-and-play in arbitrary deep learning architectures, enabling feats such as correcting miscalibrated projection orientations.