Vision-language models (VLMs) pre-trained at large scale have shown unprecedented transferability capabilities and are being progressively integrated into medical image analysis. Although its discriminative potential has been widely explored, its reliability aspect remains overlooked. This work investigates their behavior under the increasingly popular split conformal prediction (SCP) framework, which theoretically guarantees a given error level on output sets by leveraging a labeled calibration set. However, the zero-shot performance of VLMs is inherently limited, and common practice involves few-shot transfer learning pipelines, which cannot absorb the rigid exchangeability assumptions of SCP. To alleviate this issue, we propose full conformal adaptation, a novel setting for jointly adapting and conformalizing pre-trained foundation models, which operates transductively over each test data point using a few-shot adaptation set. Moreover, we complement this framework with SS-Text, a novel training-free linear probe solver for VLMs that alleviates the computational cost of such a transductive approach. We provide comprehensive experiments using 3 different modality-specialized medical VLMs and 9 adaptation tasks. Our framework requires exactly the same data as SCP, and provides consistent relative improvements of up to 27% on set efficiency while maintaining the same coverage guarantees.

This study aims to evaluate the predictive value of multiparameter characteristics of coronary computed tomography angiography (CCTA) plaque and the ratio of coronary artery volume to myocardial mass (V/M) in guiding percutaneous coronary stent implantation (PCI) in patients diagnosed with unstable angina. Patients who underwent CCTA and coronary angiography (CAG) within 2 months were retrospectively analyzed. According to CAG results, patients were divided into a medical therapy group (n=41) and a PCI revascularization group (n=37). The plaque characteristics and V/M were quantitatively evaluated. The parameters included minimum lumen area at stenosis (MLA), maximum area stenosis (MAS), maximum diameter stenosis (MDS), total plaque burden (TPB), plaque length, plaque volume, and each component volume within the plaque. Fractional flow reserve (FFR) and pericoronary fat attenuation index (FAI) were calculated based on CCTA. Artificial intelligence software was employed to compare the differences in each parameter between the 2 groups at both the vessel and plaque levels. The PCI group had higher MAS, MDS, TPB, FAI, noncalcified plaque volume and lipid plaque volume, and significantly lower V/M, MLA, and CT-derived fractional flow reserve (FFRCT). V/M, TPB, MLA, FFRCT, and FAI are important influencing factors of PCI. The combined model of MLA, FFRCT, and FAI had the largest area under the ROC curve (AUC=0.920), and had the best performance in predicting PCI. The integration of AI-derived multiparameter features from one-stop CCTA significantly enhances the accuracy of predicting PCI in angina pectoris patients, evaluating at the plaque, vessel, and patient levels.

In radiotherapy planning, acquiring both magnetic resonance (MR) and computed tomography (CT) images is crucial for comprehensive evaluation and treatment. However, simultaneous acquisition of MR and CT images is time-consuming, economically expensive, and involves ionizing radiation, which poses health risks to patients. The objective of this study is to generate CT images from radiation-free MR images using a novel quasi-supervised learning framework. In this work, we propose a quasi-supervised framework to explore the underlying relationship between unpaired MR and CT images. Normalized mutual information (NMI) is employed as a similarity metric to evaluate the correspondence between MR and CT scans. To establish optimal pairings, we compute an NMI matrix across the training set and apply the Hungarian algorithm for global matching. The resulting MR-CT pairs, along with their NMI scores, are treated as prior knowledge and integrated into the training process to guide the MR-to-CT image translation model. Experimental results indicate that the proposed method significantly outperforms existing unsupervised image synthesis methods in terms of both image quality and consistency of image features during the MR to CT image conversion process. The generated CT images show a higher degree of accuracy and fidelity to the original MR images, ensuring better preservation of anatomical details and structural integrity. This study proposes a quasi-supervised framework that converts unpaired MR and CT images into structurally consistent pseudo-pairs, providing informative priors to enhance cross-modality image synthesis. This strategy not only improves the accuracy and reliability of MR-CT conversion, but also reduces reliance on costly and scarce paired datasets. The proposed framework offers a practical 1 and scalable solution for real-world medical imaging applications, where paired annotations are often unavailable.

Predicting the final infarct after an extended time window mechanical thrombectomy (MT) is beneficial for treatment planning in acute ischemic stroke (AIS). By introducing guidance from prior knowledge, this study aims to improve the accuracy of the deep learning model for post-MT infarct prediction using pre-MT brain perfusion data. This retrospective study collected CT perfusion data at admission for AIS patients receiving MT over 6 hours after symptom onset, from January 2020 to December 2024, across three centers. Infarct on post-MT diffusion weighted imaging served as ground truth. Five Swin transformer based models were developed for post-MT infarct segmentation using pre-MT CT perfusion parameter maps: BaselineNet served as the basic model for comparative analysis, CollateralFlowNet included a collateral circulation evaluation score, InfarctProbabilityNet incorporated infarct probability mapping, ArterialTerritoryNet was guided by artery territory mapping, and UnifiedNet combined all prior knowledge sources. Model performance was evaluated using the Dice coefficient and intersection over union (IoU). A total of 221 patients with AIS were included (65.2% women) with a median age of 73 years. Baseline ischemic core based on CT perfusion threshold achieved a Dice coefficient of 0.50 and IoU of 0.33. BaselineNet improved to a Dice coefficient of 0.69 and IoU of 0.53. Compared with BaselineNet, models incorporating medical knowledge demonstrated higher performance: CollateralFlowNet (Dice coefficient 0.72, IoU 0.56), InfarctProbabilityNet (Dice coefficient 0.74, IoU 0.58), ArterialTerritoryNet (Dice coefficient 0.75, IoU 0.60), and UnifiedNet (Dice coefficient 0.82, IoU 0.71) (all P<0.05). In this study, integrating medical knowledge into deep learning models enhanced the accuracy of infarct predictions in AIS patients undergoing extended time window MT.

Kyphosis is a prevalent spinal condition where the spine curves in the sagittal plane, resulting in spine deformities. Curvature estimation provides a powerful index to assess the deformation severity of scoliosis. In current clinical diagnosis, the standard curvature estimation method for quantitatively assessing the curvature is performed by measuring the vertebral angle, which is the angle between two lines, drawn perpendicular to the upper and lower endplates of the involved vertebra. However, manual Cobb angle measurement requires considerable time and effort, along with associated problems such as interobserver and intraobserver variations. Hence, in this study, we propose UNet-based Attention Network for Thoracolumbar Vertebral Compression Fracture Angle (UANV), a vertebra angle measuring model using lateral spinal X-ray based on a deep convolutional neural network (CNN). Specifically, we considered the detailed shape of each vertebral body with an attention mechanism and then recorded each edge of each vertebra to calculate vertebrae angles.

To investigate the diagnostic capability of multiple machine learning algorithms combined with intratumoral and peritumoral ultrasound radiomics models for non-massive breast cancer in dense breast backgrounds. Manual segmentation of ultrasound images was performed to define the intratumoral region of interest (ROI), and five peritumoral ROIs were generated by extending the contours by 1 to 5 mm. A total of 851 radiomics features were extracted from these regions and filtered using statistical methods. Thirteen machine learning algorithms were employed to create radiomics models for the intratumoral and peritumoral areas. The best model was combined with clinical ultrasound predictive factors to form a joint model, which was evaluated using ROC curves, calibration curves, and decision curve analysis (DCA).Based on this model, a nomogram was developed, demonstrating high predictive performance, with C-index values of 0.982 and 0.978.The model incorporating the intratumoral and peritumoral 2 mm regions outperformed other models, indicating its effectiveness in distinguishing between benign and malignant breast lesions. This study concludes that ultrasound imaging, particularly in the intratumoral and peritumoral 2 mm regions, has significant potential for diagnosing non-massive breast cancer, and the nomogram can assist clinical decision-making.

The Segment Anything Model (SAM) was fine-tuned on the EchoNet-Dynamic dataset and evaluated on external transthoracic echocardiography (TTE) and Point-of-Care Ultrasound (POCUS) datasets from CAMUS (University Hospital of St Etienne) and Mayo Clinic (99 patients: 58 TTE, 41 POCUS). Fine-tuned SAM was superior or comparable to MedSAM. The fine-tuned SAM also outperformed EchoNet and U-Net models, demonstrating strong generalization, especially on apical 2-chamber (A2C) images (fine-tuned SAM vs. EchoNet: CAMUS-A2C: DSC 0.891 ± 0.040 vs. 0.752 ± 0.196, p < 0.0001) and POCUS (DSC 0.857 ± 0.047 vs. 0.667 ± 0.279, p < 0.0001). Additionally, SAM-enhanced workflow reduced annotation time by 50% (11.6 ± 4.5 sec vs. 5.7 ± 1.7 sec, p < 0.0001) while maintaining segmentation quality. We demonstrated an effective strategy for fine-tuning a vision foundation model for enhancing clinical workflow efficiency and supporting human-AI collaboration.

Aortic dissection (AD) is a rare condition with a high mortality rate, necessitating accurate and rapid diagnosis. This study develops an automated deep learning pipeline for identifying, segmenting, and Stanford subtyping AD using computed tomography angiography (CTA) images. This pipeline consists of four interconnected modules: aorta segmentation, AD identification, true lumen (TL) and false lumen (FL) segmentation, and Stanford subtyping. In the aorta segmentation module, a 3D full-resolution nnU-Net is trained. The segmented aorta's boundary is extracted using morphological operations and projected from multiple views in the AD identification module. AD identification is then performed using the multi-view projection data. For AD cases, a 3D nnU-Net is further trained for TL/FL segmentation based on the segmented aorta. Finally, a network is trained for Stanford subtyping using multi-view maximum density projections of the segmented TL/FL. A total of 386 CTA scans were collected for training, validation, and testing of the pipeline. For AD identification, the method achieved an accuracy of 0.979. The TL/FL segmentation for TypeA-AD and Type-B-AD achieved average Dice coefficient of 0.968 for TL and 0.971 for FL. For Stanford subtyping, the multi-view method achieved an accuracy of 0.990. The automated pipeline enables rapid and accurate identification, segmentation, and Stanford subtyping of AD using CTA images, potentially accelerating the diagnosis and treatment. The segmented aorta and TL/FL can also serve as references for physicians. The code, models, and pipeline are publicly available at https://github.com/zhuangCJ/A-pipeline-of-AD.git .

Deep convolutional neural networks (CNNs) are advancing rapidly in medical research, demonstrating promising results in diagnosis and prediction within radiology and pathology. This study evaluates the efficacy of deep learning algorithms for detecting and diagnosing dental caries using cone-beam computed tomography (CBCT) with the Mask R-CNN architecture while comparing various hyperparameters to enhance detection. A total of 2,128 CBCT images were divided into training and validation and test datasets in a 7:1:1 ratio. For the verification of tooth recognition, the data from the validation set were randomly selected for analysis. Three groups of Mask R-CNN networks were compared: A scratch-trained baseline using randomly initialized weights (R group); A transfer learning approach with models pre-trained on COCO for object detection (C group); A variant pre-trained on ImageNetfor for object detection (I group). All configurations maintained identical hyperparameter settings to ensure fair comparison. The deep learning model used ResNet-50 as the backbone network and was trained to 300epoch respectively. We assessed training loss, detection and training times, diagnostic accuracy, specificity, positive and negative predictive values, and coverage precision to compare performance across the groups. Transfer learning significantly reduced training times compared to non-transfer learning approach (p < 0.05). The average detection time for group R was 0.269 ± 0.176 s, whereas groups I (0.323 ± 0.196 s) and C (0.346 ± 0.195 s) exhibited significantly longer detection times (p < 0.05). C-group, trained for 200 epochs, achieved a mean average precision (mAP) of 81.095, outperforming all other groups. The mAP for caries recognition in group R, trained for 300 epochs, was 53.328, with detection times under 0.5 s. Overall, C-group demonstrated significantly higher average precision across all epochs (100, 200, and 300) (p < 0.05). Neural networks pre-trained with COCO transfer learning exhibit superior annotation accuracy compared to those pre-trained with ImageNet. This suggests that COCO's diverse and richly annotated images offer more relevant features for detecting dental structures and carious lesions. Furthermore, employing ResNet-50 as the backbone architecture enhances the detection of teeth and carious regions, achieving significant improvements with just 200 training epochs, potentially increasing the efficiency of clinical image interpretation.

In clinical neuroscience, the segmentation of the main white matter bundles is propaedeutic for many tasks such as pre-operative neurosurgical planning and monitoring of neuro-related diseases. Automating bundle segmentation with data-driven approaches and deep learning models has shown promising accuracy in the context of healthy individuals. The lack of large clinical datasets is preventing the translation of these results to patients. Inference on patients data with models trained on healthy population is not effective because of domain shift. This study aims to carry out an empirical analysis to investigate how transfer learning might be beneficial to overcome these limitations. For our analysis, we consider a public dataset with hundreds of individuals and a clinical dataset of glioma patients. We focus our preliminary investigation on the corticospinal tract. The results show that transfer learning might be effective in partially overcoming the domain shift.