These 609 MICCAI 2025 papers come with a code repository. Each shows an AI one-line summary below — get the verified repo link + the full 6-part summary (innovation, method, data, results, limitations) and search every MICCAI 2025 paper, free trial on arXivSub.
3D Acetabular Surface Reconstruction from 2D Pre-operative X-ray Images using SRVF Elastic Registration and Deformation Graph
Zhang, Shuai (University College London), Mazomenos, Evangelos B. (University College London)
🎯 What it does: Using SRVF elastic registration and embedded deformation maps, we reconstruct patient-specific 3D acetabular surfaces and predict suitable acetabular cup sizes based on only three 2D X-rays.
CodeObject TrackingPose EstimationSimultaneous Localization and MappingVideoMultimodality
🎯 What it does: In digital orthopedic surgery training, a visual-inertial 6D target pose tracking system utilizing dual cameras and embedded IMU is proposed, capable of real-time precise tracking of surgical tools and anatomical structures.
🎯 What it does: A method for evaluating ultrasound video plaques based on causal models is proposed, eliminating style bias caused by changes in perspective and enhancing content relevance.
A Diffusion-Driven Temporal Super-Resolution and Spatial Consistency Enhancement Framework for 4D MRI imaging
Zhou, Xuanru (Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences), Wang, Shanshan (Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences)
A flexible deep learning framework for survival analysis with medical data
Campanella, Gabriele (Icahn School of Medicine at Mount Sinai), Fuchs, Thomas J. (University of Zurich)
CodeImageTabularBiomedical DataComputed TomographyElectronic Health Records
🎯 What it does: A medical survival analysis framework based on the Deep Conditional Transformation Model (DCTM) is proposed, which can directly handle non-tabular data such as images and electronic health records.
🎯 What it does: A multi-branch few-shot cross-domain vascular segmentation framework is proposed, utilizing high-frequency auxiliary modality, dual-modal feature extraction and fusion, as well as a multi-branch feature extraction module to achieve cross-modal vascular segmentation.
A Novel ED Triage Framework Using Conditional Imputation, Multi-Scale Semantic Learning, and Cross-Modal Fusion
Xiao, Yi (Southeast University), Wang, Chunyu (Peking University People's Hospital)
CodeClassificationRecurrent Neural NetworkTransformerAuto EncoderGenerative Adversarial NetworkMultimodalityTabularBiomedical DataElectronic Health Records
🎯 What it does: This paper proposes a multimodal learning framework for triage in the emergency department, addressing the issues of missing data and sparse features.
A novel Fourier Adjacency Transformer for advanced EEG emotion recognition
Wang, Jinfeng (Kunming University of Science and Technology), Ding, Jiaman (Kunming University of Science and Technology)
CodeRecognitionTransformerTime SeriesBiomedical Data
🎯 What it does: This paper proposes a Transformer model that combines Fourier periodic features with graph structure attention—Fourier Adjacency Transformer—for EEG emotion recognition.
🎯 What it does: A set of integrated framework for percutaneous ablation of liver tumors based on 3D ultrasound has been proposed and validated, including the registration and multimodal visualization of 2D ultrasound with CT/MRI.
🎯 What it does: An unsupervised anatomical keypoint detection and Thin-Plate Spline transformation method for dMRI trajectory registration based on streamlines is proposed.
🎯 What it does: A lightweight endoscopic exposure correction network WTNet is designed, utilizing wavelet transform to separate low-frequency illumination information from high-frequency details, and modeling them through Transformer and depthwise convolution to achieve exposure correction and low-light enhancement.
🎯 What it does: A semi-supervised knowledge distillation framework is proposed, combining task-aware spatiotemporal networks (TSTNet) for left ventricle segmentation and landmark detection in echocardiography.
🎯 What it does: A multi-scale hierarchical anatomical feature fusion framework is proposed, utilizing graph neural networks and tree embedding models to achieve continuous staging and four-class classification of Alzheimer's disease and Lewy body dementia.
🎯 What it does: A unified missing modality completion model is proposed, which can generate missing modality images based on any subset of multimodal MRI.
🎯 What it does: The CLEAR framework is proposed, which achieves weakly supervised learning for 3D diagnostic tasks through frozen embeddings of 2D slices and attention aggregation, and based on this, the Lesion-enhanced Contrastive Learning (LeCL) method is developed.
🎯 What it does: This paper proposes a direct fetal birth weight prediction method based on 3D ultrasound volume, combining a multi-scale feature fusion network and a synthetic sample learning framework.
🎯 What it does: This study constructs Polyp-LDM, a latent diffusion model that combines latent space and authenticity enhancement, to generate realistic colonoscopic polyp images and corresponding masks.
🎯 What it does: Proposes the Ada-FCN framework, which utilizes adaptive cascade decomposition to learn task-related frequency bands, and constructs a unified brain functional network through frequency coupling connections, ultimately performing unified GCN for brain disease classification on this network.
🎯 What it does: A multimodal task-agnostic model MML-SurgAdapt based on CLIP is proposed, utilizing single-positive multi-label learning for phase recognition, CVS assessment, and action triplet recognition in surgical videos.
🎯 What it does: The ADAptation framework is proposed, utilizing unsupervised active learning to achieve cross-domain adaptation for breast ultrasound diagnosis, and automatically selecting the most informative target domain samples for labeling.
🎯 What it does: A dual-view collaborative training framework for dental caries detection, DVCTNet, is proposed, utilizing two perspectives of panoramic X-rays and sectioned tooth images to simulate the clinical diagnosis process and improve detection accuracy.
🎯 What it does: A framework named SelectGA has been developed, utilizing adaptive frame selection technology to process blind-scan fetal ultrasound videos for gestational age estimation.
Adaptive Stain Normalization for Cross-Domain Medical Histology
Xu, Tianyue (Johns Hopkins University), Haeffele, Benjamin D. (University of Pennsylvania)
CodeClassificationObject DetectionDomain AdaptationConvolutional Neural NetworkGenerative Adversarial NetworkImageBiomedical Data
🎯 What it does: A trainable pigment decoupling network based on the Beer-Lambert law, named BeerLaNet, is proposed and integrated as a plug-in module into target detection and classification networks to achieve adaptive pigment normalization for cross-domain pathological images.
🎯 What it does: This paper proposes Adaptively Distilled ControlNet, which uses a teacher-student dual-branch structure to perform adaptive distillation on a student model that only uses masks, guided by a teacher model that uses both images and masks during training. After training, only the student model is used to quickly generate accurately aligned medical images.
🎯 What it does: Developed an Adversarial Masked Image Modeling (AdvMIM) method for semi-supervised medical image segmentation, which constructs a masked domain and uses adversarial training to reduce the gap between the original domain and the masked domain, thereby enhancing the performance of Transformers under limited labeled data.
Analysis of Image-and-Text Uncertainty Propagation in Multimodal Large Language Models with Cardiac MR-Based Applications
Tang, Yucheng (University College London), Hu, Yipeng (King's College London)
CodeTransformerLarge Language ModelMultimodalityBiomedical DataMagnetic Resonance ImagingElectronic Health Records
🎯 What it does: A multi-modal uncertainty propagation model (MUPM) is proposed to quantify the uncertainty relationship between medical imaging and text inputs in multi-modal large language models, applied in cardiac MRI prediction tasks.
🎯 What it does: A multi-layer distillation framework based on anatomical maps (AGMD) is proposed to achieve more robust early diagnosis of Alzheimer's disease (AD) using only MRI single-modal models in the absence of multimodal data such as PET.
🎯 What it does: A few-shot ultrasound image structure detection framework TRR-CCM based on human anatomical knowledge is proposed, combining Circular Channel Mamba (CCM) and Topological Relationship Reasoning (TRR) to achieve multi-object detection;
🎯 What it does: A scale-robust hierarchical anatomical representation framework (SRHRS) based on self-supervised learning is proposed, combining multi-scale input, location scale prediction (LSP), and decomposition prediction (DP) tasks to learn consistent and hierarchical anatomical representations of chest X-ray images.
🎯 What it does: A paired CBCT→CT translation framework based on the Schrödinger bridge, named ACSB, is proposed, which achieves high-quality synthetic CT while preserving anatomical structures.
🎯 What it does: Using the medical image patch embeddings extracted by DINOv2, a Dirichlet Process Mixture Model (DPMM) is employed to model the distribution of normal samples, and the anomaly score is calculated through the cosine similarity between the embeddings and the means of the nearest mixture components, achieving unsupervised anomaly detection and segmentation.
🎯 What it does: A framework for plane detection based on anatomical constraints is proposed to address the accuracy challenges caused by boundary ambiguity and strong similarity in aortic multi-region segmentation, ultimately achieving end-to-end multi-class segmentation of the aorta.
Asynchronous Multi-Modal Learning for Dynamic Risk Monitoring of Acute Respiratory Distress Syndrome in Intensive Care Units
Feng, Yidan (Hong Kong Polytechnic University), Qin, Jing (Hong Kong Polytechnic University)
CodeAnomaly DetectionConvolutional Neural NetworkTransformerSupervised Fine-TuningMultimodalityTabularBiomedical DataElectronic Health Records
🎯 What it does: Developed a continuous ARDS risk monitoring model based on asynchronous multimodal deep learning, achieving real-time risk prediction and urgency classification for ICU patients.
🎯 What it does: A multimodal deep learning model based on attention mechanisms is proposed, utilizing whole-brain DTI microstructural indicators (FA, MD, RD, AD), structural MRI (WM, GM), and demographic features to classify upper limb motor function after stroke as good or poor.
🎯 What it does: A two-stage Attention-Guided Vector Quantized Variational Autoencoder (AG-VQ-VAE) model is proposed for brain tumor segmentation, focusing on improving the accuracy of tumor boundaries.
Augmented Reality-based Guidance with Deformable Registration in Head and Neck Tumor Resection
Yang, Qingyun (Vanderbilt University), Wu, Jie Ying (Florida International University)
CodeSegmentationPose EstimationSimultaneous Localization and MappingPoint CloudMesh
🎯 What it does: A deformable registration framework combining pre-resection surface and post-resection cavity is proposed for the re-localization of positive margins after head and neck tumor resection, and it is integrated into an augmented reality (AR) guidance system.
Automated Auditing of Upper Endoscopy Procedure Times: A Temporal Multiclass Analysis
Bravo, Diego (Universidad Nacional de Colombia), Romero, Eduardo (Universidad Nacional de Colombia)
CodeClassificationRecognitionTransformerVideoBiomedical Data
🎯 What it does: A multi-scale video sequence model MSSI based on Transformer is proposed for automatically auditing the organ and gastric region dwell time during gastroscopy.
🎯 What it does: An end-to-end framework named PolarNet is proposed for the automated segmentation of myocardial scars and characterization of their topological subtypes in polar coordinates, assisting in the screening of ventricular tachycardia (VT).
🎯 What it does: This paper proposes an unsupervised dataset shift detection framework that can automatically distinguish between prevalence shift (label shift), covariate shift, and mixed shift during the deployment of medical imaging.
🎯 What it does: This paper proposes an autoregressive-based medical image segmentation framework called AR-Seg, which can predict segmentation masks at multiple scales, achieving a segmentation process from coarse to fine.
🎯 What it does: This paper proposes a controllable adversarial diffusion model (AVDM) for generating volumes from vascular structures, addressing the issue of poor alignment of vascular structures in existing methods.
🎯 What it does: This paper proposes a graph self-supervised reconstruction fusion framework called GRAF, which jointly predicts early brain Aβ PET uptake maps in Alzheimer's disease using functional (rs-fMRI) and structural (dMRI) neural networks.
🎯 What it does: Designed and implemented BIIGMA-Net, a network that enhances the robustness of skin lesion classification through multi-head independent guided channel attention and background-invariant vector sampling.
BaMCo: Balanced Multimodal Contrastive Learning for Knowledge-Driven Medical VQA
Yazıcı, Ziya Ata (Istanbul Technical University), Ekenel, Hazım Kemal (Istanbul Technical University)
CodeClassificationRecognitionRetrievalTransformerLarge Language ModelContrastive LearningMultimodalityBiomedical Data
🎯 What it does: This paper proposes a knowledge space pre-training method based on balanced multimodal contrastive learning (BaMCo) and integrates it into large language models (LLM) to achieve medical visual question answering (Medical VQA).
Bayesian Transformers and Higher-Order Graph Matching for Cell Tracking in Serial Tissue Sections
Karami, Mostafa (University of Connecticut), Nabavi, Sheida (University of Connecticut)
CodeObject TrackingSegmentationTransformerContrastive LearningMultimodalityBiomedical Data
🎯 What it does: A novel unsupervised Bayesian Transformer combined with higher-order graph matching is proposed for cell tracking, aimed at achieving 3D tissue reconstruction in multi-channel sequential slices.
🎯 What it does: A BayeSMM framework based on a Student's t-distribution mixture model is proposed for the joint registration and segmentation of multimodal medical images.
🎯 What it does: This paper proposes BCRNet, which achieves precise detection of key points in laparoscopic liver structures through Bezier curve prediction.
🎯 What it does: A systematic retinal anomaly detection benchmark named BenchReAD has been established, which integrates multimodal data and conducts graded comparisons of methods, proposing an improved scheme called NFM-DRA based on normal feature memory.
CodeTransformerLarge Language ModelVision Language ModelImageTextMultimodalityMagnetic Resonance Imaging
🎯 What it does: A systematic evaluation of the impact of in-context learning (ICL) using visual language models on the predictive fairness for demographic subgroups such as race and gender in medical image diagnosis.
🎯 What it does: This work proposes the Bio2Vol framework, which transfers pre-trained 2D text-prompted foundational models (such as BiomedParse) to 3D medical image segmentation tasks, supporting volume-level text-prompted segmentation.
Bipartite Patient-Modality Graph Learning with Event-Conditional Modelling of Censoring for Cancer Survival Prediction
Yue, Hailin (Central South University), Wang, Jianxin (Institute of Guizhou Aerospace Measuring and Testing Technology)
CodeGraph Neural NetworkMultimodalityBiomedical Data
🎯 What it does: A bipartite graph (patient-modal) was constructed and event conditional censoring modeling (ECMC) was incorporated to integrate pathological images, genomic, and clinical data to predict the survival risk of cancer patients.
CodeCompressionExplainability and InterpretabilityGraph Neural NetworkAuto EncoderBiomedical DataMagnetic Resonance Imaging
🎯 What it does: A behavior information-based subgroup consistency brain network template (BISCoT) is proposed to compress the rs-fMRI connectivity matrix and extract interpretable sparse subnetworks.
🎯 What it does: This paper proposes a BP-nnU-Net framework that integrates blood pressure information for the segmentation of cerebral microbleeds (CMB).
🎯 What it does: A pluggable multimodal/multi-timepoint image restoration module (BME 2) is proposed to roughly estimate and recursively refine the deformation field across scans, thereby eliminating mismatches between images and improving restoration quality.
🎯 What it does: The researchers analyzed the shortcomings of existing 3D liver datasets and proposed generating high-quality synthetic liver models through HyperDiffusion and implicit neural representations (INR) to expand the dataset.
🎯 What it does: This paper proposes a medical image synthesis framework that combines registration-guided consistency with decoupled learning, suppressing misalignment noise during the training process by applying the same deformation before and after synthesis and using alignment loss.
Brain Wiring Knowledge Graph Reasoning: A Region Embedding Approach for Logical Neuronal Relation Inference
Zhou, Zhengyun (Wuhan University), Du, Bo (Wuhan University)
CodeGraph Neural NetworkGraphBiomedical Data
🎯 What it does: This paper constructs a brain connectivity graph knowledge graph from fruit flies and human cortex, and utilizes Lie group space for regional embedding to achieve logical reasoning of neuronal relationships.
🎯 What it does: A zero-shot learning framework based on brain-environment cross-attention, BECA, is proposed to achieve early diagnosis of brain network diseases using a brain-based model pre-trained with logistic regression.
🎯 What it does: This paper proposes a brain-heart-gut interactive information fusion and multi-constraint knowledge distillation framework based on whole-body PET to enhance the diagnostic accuracy of brain imaging in early Alzheimer's disease.
🎯 What it does: Proposes the BrainPrompt framework, which utilizes prompt learning to achieve domain adaptation of multi-site brain networks for distinguishing autism spectrum disorders.
BrainPrompt: Multi-Level Brain Prompt Enhancement for Neurological Condition Identification
Xu, Jiaxing (Nanyang Technological University), Feng, Mengling (National University of Singapore)
CodeClassificationGraph Neural NetworkLarge Language ModelPrompt EngineeringGraphBiomedical DataAlzheimer's Disease
🎯 What it does: A multi-layer Prompt-based brain network classification framework called BrainPrompt is proposed, integrating ROI, subject, and disease-level Prompts with GNN;
🎯 What it does: This study proposes the BREA-Depth framework, which utilizes foundational model adaptation, depth-aware CycleGAN, and airway structure-aware loss to achieve monocular depth estimation in airway endoscopy, and introduces a new structural consistency evaluation metric.
Bridging Knowledge Discrepancy in Retinal Image Analysis through Federated Multi-Task Learning
Yang, Jing (National Institute for Data Science in Health and Medicine, Xiamen University), Wang, Liansheng (Shanghai Changhai Hospital)
CodeClassificationSegmentationFederated LearningTransformerContrastive LearningImageBiomedical Data
🎯 What it does: This paper proposes the FedBKD framework, which aims to bridge the knowledge gap between global and local models in a federated multi-task learning environment for disease recognition and segmentation of retinal images using a base model.
Bridging Radiological Images and Factors with Vision-Language Model for Accurate Diagnosis of Proliferative Hepatocellular Carcinoma
Huang, Yanyan (University of Hong Kong), Yu, Lequan (University of Hong Kong)
CodeClassificationTransformerPrompt EngineeringVision Language ModelImageTextMultimodalityTabularComputed Tomography
🎯 What it does: This paper proposes the FM-Bridge method, which achieves accurate diagnosis of proliferative hepatocellular carcinoma (HCC) by textually transforming the tabulated imaging factors evaluated by clinical experts and utilizing a visual-language model (VLM) for cross-modal alignment, integrating CT images and tabular data.
CodeSegmentationConvolutional Neural NetworkImageBiomedical Data
🎯 What it does: A connectivity-based loss function called CAPE is proposed, aimed at improving the topological correctness of segmentation for curved structures (such as neurons and blood vessels) by comparing the shortest path costs between predictions and ground truths.
CardiacCLIP: Video-based CLIP Adaptation for LVEF Prediction in a Few-shot Manner
Du, Yao (Hong Kong University of Science and Technology), Li, Xiaomeng (Southern Medical University)
CodeClassificationDomain AdaptationOptimizationConvolutional Neural NetworkVision Language ModelContrastive LearningVideoTextBiomedical DataUltrasound
🎯 What it does: This paper presents CardiacCLIP, a CLIP adaptation framework for cardiac ultrasound videos, aimed at predicting left ventricular ejection fraction (LVEF) in a few-shot environment.
🎯 What it does: A framework called CardiacFlow is proposed for implicit four-chamber ventricular shape completion and generation based on flow matching, achieving efficient one-step generation and shape completion.
🎯 What it does: The CardioInterp model is proposed, utilizing a latent diffusion model and a dual-path fusion decoder to perform high-quality interpolation of B-slices in cardiovascular OCT, generating continuous and high-resolution images.
CARE-VL: A Domain-Specialized Vision-Language Model for Early ASD Screening
Yoo, Cheol-Hwan (ETRI), Jang, Jaeyoon (ETRI)
CodeClassificationRecognitionTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelVideoMultimodality
🎯 What it does: A domain-specific visual-language model named CARE-VL is proposed to identify children's social behaviors from social interaction videos and achieve early screening for autism spectrum disorder (ASD) through LLM aggregation.
🎯 What it does: A full-body PET/CT synthesis framework based on a cascaded three-dimensional diffusion model is proposed, which can generate high-fidelity 18-F FDG PET/CT volumes using only demographic variables such as age, gender, height, and weight, for virtual clinical trials and AI data augmentation.
🎯 What it does: This paper proposes a five-stage end-to-end framework that decodes visual stimuli from EEG signals and generates images similar to the original using Stable Diffusion.
🎯 What it does: A cross-domain polyp segmentation network CD-PolypNet based on the large visual model SAM has been constructed, combining internal feature distillation and dual-stream boundary focusing.
CellStyle: Improved Zero-Shot Cell Segmentation via Style Transfer
Yilmaz, Rüveyda (RWTH Aachen University), Stegmaier, Johannes (RWTH Aachen University)
CodeSegmentationDomain AdaptationConvolutional Neural NetworkDiffusion modelImageBiomedical Data
🎯 What it does: The CellStyle method is proposed, utilizing an unsupervised diffusion model for style transfer on unlabeled target data, mapping attributes such as texture and color from the target to labeled source data, generating synthetic images with original annotations for fine-tuning zero-shot cell instance segmentation models.
🎯 What it does: A brain vascular disease screening method based on color fundus photographs, CVGB-Net, is proposed, which integrates pixel-level and semantic-level features and performs discrimination through a graph network.
🎯 What it does: The CF-Seg framework is proposed, which uses a deep structural causal model to generate pseudo-healthy control images, and then employs a pre-trained U-Net to perform anatomical structure segmentation on these pseudo-healthy images, thus eliminating the need to retrain the segmentation model.
CholecMamba: A Mamba-based Multimodal Reasoning Model for Cholecystectomy Surgery
Wang, Zipei (Chinese Academy of Sciences), Dong, Di (University of Chinese Academy of Sciences)
CodeRecognitionSegmentationTransformerLarge Language ModelReinforcement LearningVideoMultimodality
🎯 What it does: CholecMamba is proposed—a multimodal reasoning model based on Mamba for comprehensive analysis and segmentation of cholecystectomy surgery videos.
Class-Conditioned Image Synthesis with Diffusion for Imbalanced Diabetic Retinopathy Grading
Zhang, Haochen (UC San Diego), An, Cheolhong (UC San Diego)
CodeGenerationData SynthesisDiffusion modelImage
🎯 What it does: This paper utilizes a conditional diffusion model to generate fundus images of diabetic retinopathy (DR) to alleviate the imbalance in training data.
ClinGRAD: Clinically-Guided Genomics and Radiomics Interpretable GNN for Dementia Diagnosis
Hassan, Salma (Mohamed Bin Zayed University of Artificial Intelligence), Yaqub, Mohammad (Mohamed bin Zayed University of Artificial Intelligence)
CodeClassificationExplainability and InterpretabilityGraph Neural NetworkMultimodalityBiomedical DataMagnetic Resonance ImagingAlzheimer's Disease
🎯 What it does: ClinGRAD, a clinical-guided heterogeneous graph neural network, was developed to diagnose Alzheimer's disease through multi-scale integration of genomics and radiomics.
CodeClassificationRetrievalConvolutional Neural NetworkVision Language ModelImageTextMultimodalityComputed TomographyElectronic Health RecordsRetrieval-Augmented Generation
🎯 What it does: A clinical data-driven retrieval-enhanced visual-language model framework is proposed for predicting the malignancy of lung nodules.
🎯 What it does: This paper proposes an end-to-end CM-DAC framework that combines YOLOv11 slice detection, multimodal contrastive pre-training, and cross-modal hierarchical fusion to achieve high-precision subtype classification of lung cancer using CT images and clinical prior information.
Clinically-guided Data Synthesis for Laryngeal Lesion Detection
Baldini, Chiara (Istituto Italiano di Tecnologia), Mattos, Leonardo S. (Istituto Italiano di Tecnologia)
CodeObject DetectionData SynthesisDiffusion modelImageBiomedical Data
🎯 What it does: Using implicit diffusion models (LDM) and ControlNet, annotated laryngoscopic images are generated based on clinical information and annotation boxes to expand medical image datasets and improve the performance of laryngeal lesion detection.
CLIP-DSA: Textual Knowledge-Guided Cerebrovascular Diseases Recognition in Multi-View Digital Subtraction Angiography
Xie, Qihang (Ningbo Institute of Materials Technology and Engineering, Chinese Academy of Sciences), Zhang, Jiong (Chinese Academy of Sciences)
CodeClassificationRecognitionConvolutional Neural NetworkVision Language ModelContrastive LearningImageMultimodalityBiomedical Data
🎯 What it does: Proposes the CLIP-DSA framework, which utilizes CLIP textual knowledge to guide the classification of cerebrovascular diseases in multi-view DSA sequences.
🎯 What it does: The ClipGS framework is proposed to achieve interactive cinematic-level volumetric medical visualization based on clipable Gaussian splatting.
🎯 What it does: Proposes the Chain-of-Cancer (CoC) framework, which utilizes three clinical modalities—genomics, methylation, and whole-slide pathology images—along with text prompts for tumor survival prediction.
CodeObject DetectionSegmentationTransformerMixture of ExpertsImageMultimodalityBiomedical DataComputed Tomography
🎯 What it does: This paper constructs the first brain CT perfusion (CTP) dataset that includes information on collateral vessels and proposes a multimodal fusion network based on the status of collateral vessels (CCGM). It extracts spatiotemporal features through a bidirectional Mamba, fuses multimodal information using a mixture of experts mechanism, and achieves two-stage stroke localization and segmentation guided by collateral priors.
🎯 What it does: This paper proposes a method for predicting the distribution of hip joint angles and estimating confidence based on ultrasound images. It utilizes UNet++ to generate heatmaps for key anatomical landmarks, and then derives the angle probability distribution and confidence indicators through Monte Carlo sampling.
🎯 What it does: A post-hoc confidence interval construction method is proposed that relies solely on binary prediction masks and does not require internal model information (such as sigmoid scores). It utilizes morphological dilation to generate nested prediction sets, ensuring a coverage rate of ≥ 1-α under a given risk α.
🎯 What it does: A semantic uncertainty quantification method based on conformal risk control, called sem-CRC, is proposed for CT image reconstruction and denoising tasks.