🎯 What it does: A weakly supervised 2D/3D image registration framework is proposed, utilizing ROI segmentation information in X-ray projections and rendered images, and employing a differentiable Dice loss to achieve joint learning and optimization of registration and segmentation.
Weighted Stratification in Multi-Label Contrastive Learning for Long-Tailed Medical Image Classification
Lin, Ying-Chih (National Yang Ming Chiao Tung University), Chen, Yong-Sheng (National Yang Ming Chiao Tung University)
CodeClassificationContrastive LearningImageBiomedical Data
🎯 What it does: This paper proposes a multi-label contrastive learning framework named ws-MulSupCon to address the long-tail distribution and comorbidity issues in medical image classification.
🎯 What it does: A fast diffusion model based on wavelet transform, WiD-PET, is proposed for the reconstruction of low-dose PET images, aimed at recovering standard dose images from ultra-low dose data.
🎯 What it does: A reinforcement learning method based on world models (TD-MPC2) is proposed and implemented for multi-task vascular navigation during mechanical thrombectomy (MT) procedures.
CodeClassificationSegmentationGenerationRetrievalTransformerLarge Language ModelAgentic AIVision Language ModelImageMultimodalityBiomedical DataBenchmark
🎯 What it does: A collaborative multi-agent system (WSI-Agents) has been constructed for the analysis of multi-modal whole slide images (WSI), integrating task allocation, internal consistency verification, external knowledge validation, and summarization modules to achieve multi-task diagnosis, report generation, and other functions.
🎯 What it does: A network called XFMamba based on Mamba is proposed for multi-view medical image classification, and a two-stage cross-view fusion mechanism is designed.
🎯 What it does: Proposed and implemented the XOCT framework, achieving deep learning translation from OCT images to OCTA images while preserving the continuity and details of blood vessels during the translation process.
🎯 What it does: A single-stage ellipse detection network, EllipseDet, is proposed to directly estimate the cardiothoracic ratio (CTR) in the fetal four-chamber view through ellipse regression, avoiding the traditional segmentation post-processing workflow.
µ2 Tokenizer: Differentiable Multi-Scale Multi-Modal Tokenizer for Radiology Report Generation
Li, Siyou (Queen Mary University of London), Zhang, Le (University of Birmingham)
CodeGenerationTransformerLarge Language ModelSupervised Fine-TuningImageTextMultimodalityComputed Tomography
🎯 What it does: A differentiable multi-scale multi-modal tokenizer µ 2 Tokenizer is proposed, combined with a large language model µ 2 LLM to achieve CT image report generation.