Multi-disease Detection in Retinal Images Guided by Disease Causal Estimation
Xie, Jianyang (University of Liverpool), Zheng, Yalin (Ningbo Institute of Materials Technology and Engineering)
CodeClassificationTransformerImage
🎯 What it does: This paper proposes a full-process multi-disease detection model based on disease causal estimation, utilizing cross-attention to obtain disease-specific features and learning a directed acyclic graph (DAG) between diseases to regularize feature learning.
Multi-modality 3D CNN Transformer for Assisting Clinical Decision in Intracerebral Hemorrhage
Xiong, Zicheng (Henan University), Yang, Fuxing (Fujian Medical University)
CodeConvolutional Neural NetworkTransformerImageMultimodalityTabularBiomedical DataComputed TomographyElectronic Health Records
🎯 What it does: A multimodal 3D CNN-Transformer model that combines CT images and clinical tabular data is proposed for surgical or conservative treatment decision-making in early-stage patients with cerebral hemorrhage.
🎯 What it does: This paper proposes a multi-scale region-aware implicit neural network for generating soft masks (alpha matte) in medical images, helping to address the resolution issues of traditional binary segmentation in fuzzy lesion areas.
Multi-stage Multi-granularity Focus-tuned Learning Paradigm for Medical HSI Segmentation
Dong, Haichuan (East China Normal University), Wang, Yan (East China Normal University)
CodeSegmentationConvolutional Neural NetworkBiomedical Data
🎯 What it does: A multi-stage, multi-granularity Focus-tuned Learning framework is proposed for medical hyperspectral image segmentation, which can simultaneously focus on subtle spectral differences at both the disease and image levels while balancing spatial-spectral features.
Multilevel Causality Learning for Multi-label Gastric Atrophy Diagnosis
Cui, Xiaoxiao (Shandong University), Li, Shuo (Harbin Institute of Technology)
CodeClassificationTransformerImageBiomedical Data
🎯 What it does: A multi-layer causal learning framework is proposed for multi-label gastric atrophy diagnosis, capable of simultaneously detecting atrophy and gastric location.
Multimodal Cross-Task Interaction for Survival Analysis in Whole Slide Pathological Images
Jiang, Songhan (Harbin Institute of Technology (Shenzhen)), Zhang, Yongbing (Harbin Institute of Technology Shenzhen)
CodeClassificationRepresentation LearningTransformerSupervised Fine-TuningImageMultimodalityBiomedical Data
🎯 What it does: A multi-modal cross-task interaction framework MCTI based on subclass classification assistance is proposed to integrate pathological WSI and gene expression for tumor survival prediction.
Multimodal Learning for Embryo Viability Prediction in Clinical IVF
Kim, Junsik (Harvard University), Pfister, Hanspeter (Harvard University)
CodeClassificationSegmentationAnomaly DetectionTransformerVideoMultimodalityTabularBiomedical DataElectronic Health Records
🎯 What it does: This paper develops a multimodal prediction model that integrates delayed embryo time-lapse videos with patients' electronic health records (EHR) to assess the viability of IVF embryos and predict pregnancy success rates.
Multimodal Variational Autoencoder for Low-cost Cardiac Hemodynamics Instability Detection
Suvon, Mohammod N. I. (University of Sheffield), Lu, Haiping (University of Sheffield)
CodeAnomaly DetectionConvolutional Neural NetworkAuto EncoderMultimodalityBiomedical DataElectronic Health RecordsElectrocardiogram
🎯 What it does: This paper proposes a multimodal variational autoencoder (CardioVAE X,G) that utilizes low-cost chest X-ray and electrocardiogram data for predicting cardiovascular hemodynamic instability (CHDI), particularly the estimation of pulmonary artery wedge pressure (PAWP);
Myocardial Scar Enhancement in LGE Cardiac MRI using Localized Diffusion
Hasny, Marta (Beth Israel Deaconess Medical Center and Harvard Medical School), Nezafat, Reza (Beth Israel Deaconess Medical Center and Harvard Medical School)
🎯 What it does: This study proposes a local conditional diffusion model to enhance the contrast between scars and normal myocardium in LGE cardiac magnetic resonance imaging.
Neural Cellular Automata for Lightweight, Robust and Explainable Classification of White Blood Cell Images
Deutges, Michael (Helmholtz Zentrum München - German Research Center for Environmental Health), Marr, Carsten (Helmholtz Zentrum München - German Research Center for Environmental Health)
CodeClassificationExplainability and InterpretabilityImage
🎯 What it does: A lightweight, robust, and interpretable single-cell leukocyte classification method based on Neural Cellular Automata (NCA) is proposed and validated on three datasets from different sources.
CodeRetrievalTransformerLarge Language ModelContrastive LearningTextBiomedical Data
🎯 What it does: NeuroConText is proposed, a brain-text mapping framework based on contrastive learning that can associate the text of neuroscience articles with brain activation coordinates and generate brain activation maps from text.
🎯 What it does: NeuroLink is proposed, a 3D neuron segmentation and reconstruction method that combines dynamic snake convolution, weighted local connectivity loss, and multi-task learning, addressing the weak signal problem in low-contrast images.
🎯 What it does: Through systematic evaluation, validation trap identification, and comparative analysis, the current state of 3D medical image segmentation is re-examined, and more rigorous validation standards are proposed.
🎯 What it does: The paper systematically evaluates the performance of diffusion models in medical image denoising, particularly the challenges in maintaining the integrity of image content.
🎯 What it does: This paper studies a 3D medical image sequence regression framework based on Neural Ordinary Differential Equations (NODER) to predict missing or future time point images.
🎯 What it does: A noise level adaptive diffusion model (Nila-DC) is proposed to accelerate MRI image reconstruction, capable of maintaining high-quality reconstruction in the presence of measurement noise.
🎯 What it does: This paper proposes an unsupervised image registration method for preoperative and follow-up MRI of brain tumors, which improves registration accuracy by generating and utilizing a noise-reduced structural inconsistency activation map (NR-IAM) to locate and correct structurally inconsistent areas.
🎯 What it does: A non-adversarial generative framework based on VQ-VAE and dynamic Seq2Seq (VQ-Seq2Seq) is proposed, which constructs a common latent representation for multi-sequence MRI using a discretized latent space, thereby enabling the synthesis of missing sequences.
🎯 What it does: A framework for freehand ultrasound (US) three-dimensional reconstruction without external trackers is proposed, jointly estimating the rigid motion of the probe and the non-rigid deformation of soft tissues.
🎯 What it does: This paper proposes an unsupervised acoustic mapping method that sonifies the epiretinal membrane (ERM) and the slight elevation of the retinal structure through real-time OCT A-scans, assisting surgeons in identifying suitable starting points during peeling surgery.
🎯 What it does: A novel On-the-Fly Guidance (OFG) training framework is proposed, which utilizes a differentiable optimizer to generate pseudo-labels in real-time during the training process, elevating the learning-based image registration model from unsupervised to near-supervised status.
🎯 What it does: This paper proposes to reformulate image registration as a two-stage multi-class segmentation problem that seeks corresponding ROIs in two images, and accomplishes registration in an unsupervised manner.
🎯 What it does: Developed an online dense 3D reconstruction and tracking framework suitable for binocular endoscopic videos, capable of real-time construction and tracking of soft tissue surfaces.
🎯 What it does: This paper proposes a differential motion model based on a low-dimensional linear Gaussian state space model, capable of estimating and predicting motion in continuous medical image sequences.
Open-Set Semi-Supervised Medical Image Classification with Learnable Prototypes and Outlier Filter
He, Along (Nankai University), Fu, Huazhu (Agency for Science, Technology and Research)
CodeClassificationAnomaly DetectionTransformerSupervised Fine-TuningImageBiomedical Data
🎯 What it does: Proposes the OpenSSC framework for open-set semi-supervised classification of medical images, which includes learnable prototypes, multiple binary discriminators, and a joint anomaly filter.
ORacle: Large Vision-Language Models for Knowledge-Guided Holistic OR Domain Modeling
Özsoy, Ege (Technische Universität München), Navab, Nassir (Technische Universität München)
CodeSegmentationGenerationData SynthesisTransformerLarge Language ModelVision Language ModelDiffusion modelImageMultimodality
🎯 What it does: A system called ORacle based on a large visual language model (LVLM) is proposed, capable of generating semantic scene graphs of operating rooms directly from multi-view RGB images, achieving global operating room modeling.
ORCGT: Ollivier-Ricci Curvature-based Graph Model for Lung STAS Prediction
Cen, Min (University of Science and Technology of China), Wang, Liansheng (Shanghai Changhai Hospital)
CodeClassificationSegmentationGraph Neural NetworkImageBiomedical Data
🎯 What it does: This paper proposes a graph model based on Ollivier-Ricci curvature, which first extracts circular regions from rough tumor edge annotations and then uses graph neural networks to model the cells within that region, thereby automatically predicting STAS in lung adenocarcinoma slices.
🎯 What it does: A two-stage knowledge distillation framework is proposed, combining real-label-based Overlay edge cropping and stitching data augmentation for semi-supervised medical image segmentation, particularly targeting left atrium (LA) MRI data.
🎯 What it does: A semi-supervised medical image segmentation method based on pair shuffling is proposed, which splits labeled and unlabeled images into small patches and randomly rearranges them during training to enhance local information learning.
Liu, Jiashuai (Xi'an Jiaotong University), Gao, Zeyu (University of Cambridge)
CodeClassificationExplainability and InterpretabilityImage
🎯 What it does: A Prototype Attention-based Multiple Instance Learning (PAMIL) model is proposed for multi-label classification of whole slide images (WSI), providing instance-level and prototype-level explanations through a dual-branch prototype and attention mechanism.
🎯 What it does: Under various PET scanners, a method called PETITE is proposed to achieve the reconstruction from short-term PET scans to long-term scans using Parameter-Efficient Fine-Tuning (PEFT);
🎯 What it does: The PASTA framework is proposed, utilizing conditional diffusion models to convert structural magnetic resonance imaging (MRI) into functional positron emission tomography (PET) images, generating synthetic PET that retains pathological information related to Alzheimer's disease.
Patch-Slide Discriminative Joint Learning for Weakly-Supervised Whole Slide Image Representation and Classification
Yu, Jiahui (Zhejiang University), Xu, Yingke (University of British Columbia)
CodeClassificationRepresentation LearningTransformerContrastive LearningImageBiomedical Data
🎯 What it does: The PSJA-MIL framework is proposed, which jointly optimizes patch and whole slide image (WSI) level features. It enhances the discriminative power of weakly supervised WSI classification through patch contrastive loss estimation, adaptive cross-entropy loss, and a prototype cosine similarity-based PCS-classifier.
🎯 What it does: A weakly supervised multi-class pathological image segmentation framework called PathMamba is proposed, which can learn segmentation masks from pixel-level and block-level features using only image-level labels.
Pathological Semantics-Preserving Learning for H&E-to-IHC Virtual Staining
Chen, Fuqiang (Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences), Qin, Wenjian (Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences)
🎯 What it does: A new H&E-to-IHC virtual staining method called PSPStain is proposed, which can directly generate IHC images from H&E images while preserving molecular-level protein expression information and addressing spatial mismatch issues.
PathoTune: Adapting Visual Foundation Model to Pathological Specialists
Lu, Jiaxuan, Zhang, Shaoting (Zhongda Hospital)
CodeDomain AdaptationTransformerPrompt EngineeringMultimodalityBiomedical Data
🎯 What it does: A multi-modal prompt tuning framework named PathoTune is proposed to quickly transfer visual or pathological base models to various pathological diagnosis tasks.
🎯 What it does: This study presents the first patient-specific, real-time, non-tracking brain ultrasound segmentation framework, which generates virtual ultrasound scans using preoperative MRI and synthesizes intraoperative ultrasound (iUS), followed by training a 2D UNet on the synthesized data for brain tumor target segmentation.
PEPSI: Pathology-Enhanced Pulse-Sequence-Invariant Representations for Brain MRI
Liu, Peirong (Harvard Medical School and Massachusetts General Hospital), Iglesias, Juan E. (Harvard Medical School and Massachusetts General Hospital)
🎯 What it does: We propose PEPSI, a pulse sequence invariant feature learning model based on synthetic lesion encoding, for brain MRI image synthesis and lesion segmentation.
🎯 What it does: A Perspective+Unet framework is proposed, combining dual-path encoding, efficient non-local Transformer blocks, and cross-scale fusion to enhance 3D medical image segmentation.
🎯 What it does: A multi-task model based on physical priors is proposed, utilizing non-contrast CT images to achieve segmentation of aortic dissection, synthetic contrast CT generation, and classification diagnosis.
Physics informed neural networks for estimation of tissue properties from multi-echo configuration state MRI
Adams-Tew, Samuel I. (University of Utah), Joshi, Sarang (University of Utah)
CodeBiomedical DataMagnetic Resonance ImagingPhysics Related
🎯 What it does: Develop a simulation of imaging signals for configuration states based on physical modeling, and use a fully connected neural network (PINN) to estimate T1, T2, T2*, and flip angle from multi-echo data in one go; evaluate the performance of different normalization methods and network structures in simulations and salt pork experiments.
🎯 What it does: A deep learning motion correction method based on physical information, PHIMO, was designed and implemented for the automatic detection and exclusion of motion-affected k-space lines in quantitative MRI (T2* estimation).
PitVQA: Image-grounded Text Embedding LLM for Visual Question Answering in Pituitary Surgery
He, Runlong (University College London), Islam, Mobarakol (University College London)
CodeTransformerLarge Language ModelVision Language ModelImageVideoText
🎯 What it does: The PitVQA dataset and the PitVQA-Net model are proposed for the visual question answering task in endonasal transsphenoidal pituitary surgery.
🎯 What it does: This paper proposes a method for estimating Gleason grading of prostate cancer based on multimodal MRI, with the core being the Poisson Ordinal Network (PON), which enables direct prediction of grading from images.
🎯 What it does: This paper proposes a multimodal deep learning pipeline that utilizes geometric encoding and clinical context to predict shape changes in the affected femur from the normal side of the patient's femur.
Prior Activation Map Guided Cervical OCT Image Classification
Wang, Qingbin (Wuhan University), Ma, Yutao (Central China Normal University)
CodeClassificationExplainability and InterpretabilityConvolutional Neural NetworkImage
🎯 What it does: By constructing a prior activation map (PAM) and aligning it with the model-generated CAM during training, high-risk lesion classification of cervical OCT images is achieved.
🎯 What it does: This paper presents PRISM, a 3D medical image interactive segmentation model that supports multiple visual prompts and can continuously improve segmentation results over multiple iterations.
🎯 What it does: A progressive growing of patch size implicit curriculum learning strategy is proposed and integrated into the nnU-Net framework for medical image segmentation training.
🎯 What it does: For single-source domain generalization (SDG) in medical image segmentation, a domain-adaptive prompt framework DAPSAM based on the large model Segment Anything Model (SAM) is proposed, which combines a universal adapter and a prototype-based prompt generator.
CodeObject DetectionAnomaly DetectionTransformerSupervised Fine-TuningPrompt EngineeringVision Language ModelImage
🎯 What it does: By prompting a dental perspective-language model, tooth numbering is performed using the dental symbol system (FDI) to achieve multi-level tooth anomaly detection.
Prompting Whole Slide Image Based Genetic Biomarker Prediction
Zhang, Ling (DAMO Academy, Alibaba Group), Wang, Yan (East China Normal University)
CodeClassificationSegmentationExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageBiomedical Data
🎯 What it does: The PromptBio framework is proposed, utilizing pathology text prompts generated by large language models and a visual language model to achieve foreground instance selection, fine-grained pathological component clustering, and interactive modeling, predicting genetic biomarkers such as MSI and BRAF from H&E slides.
PromptSmooth: Certifying Robustness of Medical Vision-Language Models via Prompt Learning
Hussein, Noor (Mohamed Bin Zayed University of Artificial Intelligence), Nandakumar, Karthik (Mohamed Bin Zayed University of Artificial Intelligence)
CodeClassificationPrompt EngineeringVision Language ModelImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: Proposes the PromptSmooth method, which utilizes prompt learning to achieve certified robustness for pre-trained medical vision-language models without the need for retraining;
🎯 What it does: A probabilistic registration method PULPo is proposed, which can perform unsupervised registration through hierarchical variational inference and Laplacian pyramid model deformation fields.
🎯 What it does: A dynamic brain network analysis method based on Quality-Aware Fuzzy Min-Max Neural Network (QFMMNet) is proposed, which extracts dynamic functional connectivity features using multi-view learning, processes fuzzy information with multi-view fuzzy min-max neural networks, and considers the quality uncertainty of different windows through a quality-aware integration module to achieve classification of multi-window dynamic functional connectivity.
🎯 What it does: Under extremely low data conditions, few-shot medical video object segmentation is performed using existing medical image annotations and sparse video annotations;
🎯 What it does: A no-reference axial super-resolution method based on implicit neural representation (INR) and two-dimensional diffusion model priors is proposed to restore isotropic resolution in 3D microscopy images.
Refining Intraocular Lens Power Calculation: A Multi-modal Framework Using Cross-layer Attention and Effective Channel Attention
Zhou, Qian, Wang, Yong (Wuhan University)
CodeConvolutional Neural NetworkImageMultimodalityBiomedical Data
🎯 What it does: This study proposes a multimodal deep learning framework for predicting IOL optical power, integrating OCT images and biological features, and achieving accurate predictions through cross-layer attention and channel attention.
Region-Specific Retrieval Augmentation for Longitudinal Visual Question Answering: A Mix-and-Match Paradigm
Yung, Ka-Wai (University College London), Mazomenos, Evangelos B. (University College London)
CodeRecognitionRetrievalRecurrent Neural NetworkContrastive LearningImageMultimodalityBiomedical DataElectronic Health RecordsBenchmarkRetrieval-Augmented Generation
🎯 What it does: Designed and evaluated RegioMix, a region-level retrieval-enhanced longitudinal visual question answering framework that utilizes mixed matching to generate pseudo-difference descriptions and aligns the differences of input pairs with questions through dual alignment, thereby answering differential questions about chest X-ray images.
Reliable Multi-View Learning with Conformal Prediction for Aortic Stenosis Classification in Echocardiography
Gu, Ang Nan (University of British Columbia), Abolmaesumi, Purang (University of British Columbia)
CodeClassificationBiomedical DataUltrasound
🎯 What it does: A training method named RT4U is proposed, aimed at improving the classification of aortic stenosis in echocardiography by introducing uncertainty, especially when dealing with imperfect data.
🎯 What it does: This paper proposes a Reliable Source Approximation (RSA) method to achieve unsupervised adaptation of source domain models in the segmentation task of vestibular schwannoma MRI without source data.
Representation Learning with a Transformer-Based Detection Model for Localized Chest X-Ray Disease and Progression Detection
Eshraghi Dehaghani, Mehrdad (McMaster University), Moradi, Mehdi (University of Illinois Urbana-Champaign)
CodeClassificationObject DetectionRepresentation LearningConvolutional Neural NetworkTransformerImageBiomedical Data
🎯 What it does: Utilizing a Transformer-based detection model (DETR) to locate anatomical regions in chest X-ray images, and extracting feature vectors from this detection model, followed by local disease detection and monitoring of local disease progression in the same feature space.
🎯 What it does: Proposed the 'Detour Connectivity' metric to capture the indirect relationship between functional connectivity (FC) and structural connectivity (SC), and based on this, designed the SC-FC Detour Network (SFDN) for graph neural network modeling.
CodeDomain AdaptationKnowledge DistillationConvolutional Neural NetworkImageBiomedical Data
🎯 What it does: The Reprogramming Distillation (RD) framework is proposed, which maps the feature space of medical foundation models to the distribution of downstream tasks through training reprogrammable modules, and then co-trains with a lightweight student model while sharing classifiers to achieve efficient downstream transfer.
Resolving Variable Respiratory Motion From Unsorted 4D Computed Tomography
Huang, Yuliang (University College London), McClelland, Jamie R. (University College London)
CodeOptimizationImageComputed Tomography
🎯 What it does: This paper proposes a 4DCT motion model based on hyper-gradient optimization that can estimate and compensate for motion errors caused by irregular breathing from unordered CT segments without external respiratory monitoring signals.
RET-CLIP: A Retinal Image Foundation Model Pre-trained with Clinical Diagnostic Reports
Du, Jiawei (Beijing Institute of Technology), Wang, Ningli (Beijing Institute of Ophthalmology)
CodeClassificationRepresentation LearningTransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: This paper proposes RET-CLIP, an audiovisual language foundation model based on CLIP, which is pre-trained using 193,865 clinical retinal images and diagnostic reports.
🎯 What it does: This paper constructs an abdominal multi-organ segmentation dataset named RAOS and establishes a robustness evaluation benchmark on this dataset.
🎯 What it does: Through information theory analysis, this paper clarifies the theoretical foundation of autoencoders in medical anomaly detection and proposes a strategy to achieve optimal performance by utilizing latent space entropy constraints.
🎯 What it does: A two-stage network architecture is proposed to decouple cell counting and localization. First, a counter network estimates the total count and generates a coarse density map based on intermediate features, and then a locator network generates a high-resolution cell position map. A global information propagation module is introduced to enhance counting accuracy.
🎯 What it does: This study proposes a cloud-based whole-slide digital workflow that utilizes low-cost microscope-captured 10X videos to automatically stitch, deblur, and reconstruct super-resolution images at 40X quality.
🎯 What it does: Proposes RetMIL, a multi-instance learning framework based on a retention mechanism for the classification task of whole slide images (WSI).
🎯 What it does: A Dependency Merging Attention (DMA) module is proposed, which dynamically aggregates similar feature tokens into prototypes, reducing redundant dependencies in ViT self-attention;
🎯 What it does: A RIP-AV framework that combines Representative Instance Pre-training (RIP) with Context-Aware Networks (PCF, DA) is proposed for high-precision retinal artery-vein segmentation.
🎯 What it does: The study uses Weighted Conformal Prediction (WCP) in three-dimensional medical image segmentation to estimate the prediction interval of lesion volume, addressing the issue of covariate shift between the calibration set and the test set.
🎯 What it does: A semi-supervised multimodal medical image segmentation framework is proposed, which enhances segmentation performance in scenarios with scarce annotations and modality mismatch by combining cross-modal collaboration and contrastive consistency learning.
🎯 What it does: This paper proposes a deep learning framework that combines the Iterative Compressed Feature Decoupling Network (ICFDNet) with category accuracy adaptive adversarial training for fine-grained classification of fatty liver ultrasound images, significantly improving the model's recognition balance across different categories.
🎯 What it does: A method named S-SAM is proposed, which achieves precise segmentation of medical images by performing singular value transformation only on the weight matrix of the image encoder of the Segment Anything Model (SAM), using category names as text prompts.
SALI: Short-term Alignment and Long-term Interaction Network for Colonoscopy Video Polyp Segmentation
Hu, Qiang (Huazhong University of Science and Technology), Wang, Zhiwei (Huazhong University of Science and Technology)
CodeSegmentationConvolutional Neural NetworkVideo
🎯 What it does: A network called SALI, which combines short-term alignment and long-term interaction, is designed for real-time segmentation of polyps in endoscopic videos.
🎯 What it does: A MR image segmentation network named SANGRE is proposed, which combines a Transformer encoder with two new modules to achieve fine segmentation.
SaSaMIM: Synthetic Anatomical Semantics-Aware Masked Image Modeling for Colon Tumor Segmentation in Non-contrast Abdominal Computed Tomography
Dai, Pengyu (Institute of Integrated Research, Institute of Science Tokyo), Suzuki, Kenji (Institute of Integrated Research, Institute of Science Tokyo)
🎯 What it does: This paper proposes SaSaMIM, a task-oriented self-supervised pre-training framework for colon tumor segmentation. It enhances early tumor segmentation performance on non-contrast CT by synthesizing intestinal and tumor semantics and performing masked image reconstruction under a dual-branch decoder in both frequency and spatial domains.
🎯 What it does: A proactive learning method based on structure and boundary consistency, SBC-AL, is proposed to enhance medical image segmentation results using Structure-aware Feature Prediction (SFP) and Attention Segmentation Refinement (ASR) modules, with sample queries conducted based on Structure Consistency Score (SCS) and Boundary Consistency Score (BCS).
SCMIL: Sparse Context-aware Multiple Instance Learning for Predicting Cancer Survival Probability Distribution in Whole Slide Images
Yang, Zekang (Chinese Academy of Sciences), Wang, Xiangdong (Chinese Academy of Sciences)
CodeClassificationTransformerImageBiomedical Data
🎯 What it does: A sparse context-aware multi-instance learning framework (SCMIL) is proposed to predict the cancer survival probability distribution of patients from whole slide images;
🎯 What it does: In semi-supervised medical image segmentation, the authors propose the SDCL framework, which utilizes the segmentation differences of two student models to achieve self-correcting learning.
🎯 What it does: A neural surface reconstruction method based on SDF, called SDFPlane, is designed for the rapid and high-precision reconstruction of deformable soft tissues from binocular endoscopic videos.
🎯 What it does: SegMamba is proposed, a 3D medical image segmentation framework based on Mamba, utilizing three-directional Mamba (ToM) and gated spatial convolution (GSC) to achieve full-scale long-range modeling.
SegNeuron: 3D Neuron Instance Segmentation in Any EM Volume with a Generalist Model
Zhang, Yanchao (Chinese Academy of Sciences), Han, Hua (Chinese Academy of Sciences)
CodeSegmentationTransformerSupervised Fine-TuningMultimodalityBiomedical Data
🎯 What it does: Construct a large-scale multi-resolution multimodal multi-species EM database EMNeuron, and based on this, train a general 3D neuron instance segmentation model SegNeuron to achieve zero-shot segmentation of any EM volume;
🎯 What it does: A framework called SPSS based on self-accelerated sample selection is proposed for extremely low-supervision medical image segmentation, aiming to alleviate the label scarcity problem by improving the quality of pseudo-labels generated by registration.
🎯 What it does: A self-supervised 3D vascular skeleton completion method is proposed, which uses a synthetic skeleton breakpoint to train the network to reconnect broken skeletons.
Self-Supervised Contrastive Graph Views for Learning Neuron-level Circuit Network
Li, Junchi (Wuhan University), Du, Bo (Wuhan University)
CodeClassificationRepresentation LearningGraph Neural NetworkContrastive LearningGraphBiomedical Data
🎯 What it does: This paper proposes a self-supervised contrastive learning framework, FlyGCL, for learning topological features from neuron-level circuit networks, and utilizes these features for neuron classification and connectivity prediction.
🎯 What it does: A blind-slice self-supervised learning framework called SOAD is proposed for the one-time removal of noise and bulk motion artifacts (BMA) from awake OCTA volumes while preserving capillary details.
🎯 What it does: This paper proposes a self-supervised k-space regularization method called PISCO, which is combined with neural implicit k-space representation (NIK) to achieve motion-resolved reconstruction of dynamic abdominal MRI.
🎯 What it does: This paper proposes a self-regularized UNet that combines semantic consistency regularization and internal feature distillation losses to address the issues of supervision imbalance and feature redundancy in medical image segmentation.
Semantics-Aware Attention Guidance for Diagnosing Whole Slide Images
Liu, Kechun (University of Washington), Shapiro, Linda G. (UCLA)
CodeClassificationSegmentationExplainability and InterpretabilityConvolutional Neural NetworkTransformerContrastive LearningImage
🎯 What it does: Proposes the Semantics-Aware Attention Guidance (SAG) framework, which converts diagnostic entities into attention signals and uses flexible attention loss to guide whole slide image (WSI) diagnostic models to better focus on pathologically important areas.
🎯 What it does: This paper proposes a semi-supervised contrastive VAE (SS-cVAE) for digital pathology images to achieve separable interpretable latent spaces for detecting tumor-infiltrating lymphocytes (TIL).
Semi-Supervised Learning for Deep Causal Generative Models
Ibrahim, Yasin (University of Oxford), Kamnitsas, Konstantinos (University of Birmingham)
CodeGenerationData SynthesisExplainability and InterpretabilityAuto EncoderImageBiomedical DataMagnetic Resonance ImagingElectronic Health Records
🎯 What it does: A semi-supervised deep causal generative model is proposed, capable of training with fully labeled, partially labeled, and unlabeled data, and generating realistic counterfactual images for missing labeled samples.
Semi-supervised Segmentation through Rival Networks Collaboration with Saliency Map in Diabetic Retinopathy
Kim, Eunjin (VUNO Inc.), Park, Hyunjin (VUNO Inc.)
CodeSegmentationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a semi-supervised multi-lesion segmentation method named RiCo, which achieves pixel-level segmentation of diabetic retinopathy through a competitive network and saliency maps.