π― What it does: This paper proposes the DECO framework, which achieves feature-level cross-domain data augmentation by decoupling retinal image features into DR-related semantic features and domain noise, and trains a DR grading model with better domain generalization capabilities based on this.
CodeGenerationData SynthesisDiffusion modelImageBiomedical Data
π― What it does: An Adaptive Deep Control Diffusion Network (ADD) is proposed, capable of generating progressive samples of pathological images under very few sample conditions, for data augmentation and improving downstream classification performance.
CodeDomain AdaptationRepresentation LearningContrastive LearningImageBiomedical Data
π― What it does: A gene-guided Siamese representation learning framework called PathoTME is proposed, which uses panoramic whole slide images (WSI) of all cancer types to predict tumor microenvironment (TME) subtypes;
π― What it does: This study investigates an automatic segmentation method for residual glioblastoma in early postoperative magnetic resonance imaging (EPMR), proposing improved sampling strategies and network structures to enhance segmentation and classification performance, and comparing it with manual annotations.
GMoD: Graph-driven Momentum Distillation Framework with Active Perception of Disease Severity for Radiology Report Generation
Xiang, ZhiPeng, Zhang, Liqiang (Huazhong University Of Science And Technology)
CodeGenerationKnowledge DistillationGraph Neural NetworkTransformerImageTextBiomedical Data
π― What it does: This paper proposes a graph-driven momentum distillation framework GMoD, which achieves active perception of disease severity in X-ray images through graph attention and momentum distillation, thereby generating more accurate and semantically consistent radiology reports.
π― What it does: A gradient-guided co-retention feature pyramid network (G2CR-FPN) is proposed for low-dose CT image denoising, which integrates multi-scale features and gradient enhancement to improve noise suppression and detail preservation.
π― What it does: To address the registration challenges of cardiac diffusion tensor CMR images under respiratory and cardiac motion as well as low signal-to-noise ratio conditions, an inter-group deformation registration framework based on implicit templates is proposed. Motion correction and text feature preservation are achieved through tensor embedding to generate pseudo-images and a differentiable mutual information loss.
HAMIL-QA: Hierarchical Approach to Multiple Instance Learning for Atrial LGE MRI Quality Assessment
Sultan, K. M. Arefeen (University of Utah), Elhabian, Shireen Y. (University of Utah)
CodeClassificationExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationConvolutional Neural NetworkImageBiomedical DataMagnetic Resonance Imaging
π― What it does: A hierarchical multi-instance learning framework named HAMIL-QA has been developed for the automatic assessment of the diagnostic quality of left atrial Late Gadolinium Enhancement (LGE) MRI scans.
π― What it does: This paper proposes a framework that combines hard negative sample mining with multi-instance ranking loss to improve the classification of whole slide images (WSI).
HATs: Hierarchical Adaptive Taxonomy Segmentation for Panoramic Pathology Image Analysis
Deng, Ruining, Huo, Yuankai (Vanderbilt University Medical Center)
CodeSegmentationTransformerBiomedical Data
π― What it does: A panoramic kidney tissue segmentation method based on hierarchical adaptive taxonomy, HATs, is proposed, capable of segmenting 15 structures from regional to cellular levels simultaneously.
π― What it does: A systematic evaluation of five open-source Hausdorff distance implementations is conducted, comparing their differences with a grid reference and discussing the impact of implementation differences on the results.
π― What it does: This paper proposes HecVL, a model that utilizes hierarchical video-text pairs for video-language pre-training, aimed at zero-shot surgical phase recognition.
π― What it does: A multi-prototype network driven by hemodynamics (HDMPNet) is proposed to achieve single-instance segmentation of breast DCE-MRI tumors, capable of handling diverse tumor sizes, shapes, and multi-connected regions.
HiA: Towards Chinese Multimodal LLMs for Comparative High-Resolution Joint Diagnosis
Ding, Xinpeng (Hong Kong University of Science and Technology), Li, Xiaomeng (Southern Medical University)
CodeRecognitionSegmentationData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelMultimodalityBiomedical DataMagnetic Resonance ImagingComputed Tomography
π― What it does: This paper proposes a high-resolution instruction-aware adapter HiA and a local Chinese medical multimodal dataset Chili-Joint, addressing the limitations of existing Chinese medical MLLMs in terms of single images, low resolution, and translation biases.
π― What it does: This study investigates cognitive score prediction based on functional MRI and proposes the SW-HGL framework, which includes hierarchical graph learning and small-world brain connectivity graphs.
π― What it does: A hierarchical multi-instance learning (H-MIL) framework is proposed, incorporating relative specific similarity (RSS) loss in COPD grading to achieve more refined classification of lung diseases.
π― What it does: A hierarchical symmetric normalization registration framework (HSyN) is proposed, which estimates bidirectional half displacement fields through a symmetric normalization network and learns the inverse displacement using a deformation inverse network, achieving high-precision medical image registration under large deformations.
Hierarchical Text-to-Vision Self Supervised Alignment for Improved Histopathology Representation Learning
Watawana, Hasindri (Mohamed Bin Zayed University of Artificial Intelligence), Khan, Fahad Shahbaz (Mohamed Bin Zayed University of Artificial Intelligence)
CodeClassificationRepresentation LearningConvolutional Neural NetworkLarge Language ModelVision Language ModelContrastive LearningImageBiomedical Data
π― What it does: By combining automatically generated hierarchical natural language descriptions with visual hierarchies, a Hierarchical Language-tied Self-Supervision (HLSS) framework is proposed to achieve hierarchical alignment in self-supervised learning of medical images (histology).
High-resolution Medical Image Translation via Patch Alignment-Based Bidirectional Contrastive Learning
Zhang, Wei (City University of Hong Kong), Li, Xinyue (City University of Hong Kong)
CodeImage TranslationGenerationData SynthesisGenerative Adversarial NetworkContrastive LearningImageBiomedical Data
π― What it does: A bidirectional contrastive learning model based on Patch alignment (PPT) is proposed, which can quickly generate corresponding IHC stained images from H&E stained images in seconds.
Yin, Chong (Hong Kong Baptist University), Yuen, Pong C. (Chinese University of Hong Kong)
CodeSegmentationGenerationData SynthesisSupervised Fine-TuningDiffusion modelAuto EncoderImageBiomedical Data
π― What it does: A text prompt based on spatial and morphological attributes is proposed to guide diffusion models in generating high-quality pathological images with diagnostic features, and to achieve a quantitative assessment of the quality of synthetic images.
π― What it does: A hierarchical multi-organ graph network (HoG-Net) was constructed to transform the primary tumors and organs at risk (OAR) annotated by doctors in CT images into a graph structure, and to predict local recurrence in patients with head and neck squamous cell carcinoma (HNSCC) using a graph attention network.
Ridzuan, Muhammad (Mohamed Bin Zayed University of Artificial Intelligence), Yaqub, Mohammad (Mohamed bin Zayed University of Artificial Intelligence)
CodeClassificationExplainability and InterpretabilityConvolutional Neural NetworkTransformerBiomedical DataComputed TomographyElectronic Health Records
π― What it does: The HuLP (Human-in-the-Loop for Prognosis) model is proposed and implemented, supporting clinical doctors in manually intervening in model predictions during the inference phase, while learning from imaging data to fill in missing clinical variables and complete survival time predictions.
π― What it does: A generative self-supervised pre-training method called HySparK based on a hybrid CNN-Transformer architecture has been developed, implementing self-supervised learning for 3D medical images using a bottom-up sparse mask and hierarchical decoding.
π― What it does: This paper proposes I Net, which utilizes implicitly parameterized INR to dynamically generate model parameters, better addressing the contextual alignment inconsistency problem in medical image segmentation.
π― What it does: The IarCAC model is proposed, utilizing instance-aware sparse attention Transformer and Fourier domain guidance module to achieve coronary artery calcification segmentation.
π― What it does: This paper proposes an instantiation motion correction pipeline (IM-MoCo) based on motion-guided implicit neural representation (INR), which effectively suppresses MRI motion artifacts while preserving anatomical structures.
Image Distillation for Safe Data Sharing in Histopathology
Li, Zhe (Friedrich-Alexander-UniversitΓ€t Erlangen-NΓΌrnberg), Kainz, Bernhard (Imperial College London)
CodeClassificationData SynthesisKnowledge DistillationConvolutional Neural NetworkDiffusion modelContrastive LearningImageBiomedical Data
π― What it does: This paper proposes InfoDist, which utilizes latent diffusion models to generate readable synthetic images. It selects the samples with the highest information content through Infomap community detection, constructing a small-scale shareable synthetic dataset for downstream classification.
π― What it does: This paper proposes a Deep Geometric Correspondence Implicit Network (DGCI) that utilizes implicit neural networks for high-fidelity modeling of the lung airway tree structure in continuous space, achieving skeletonization and fracture repair.
Improved Classification Learning from Highly Imbalanced Multi-Label Datasets of Inflamed Joints in [99mTc]Maraciclatide Imaging of Arthritic Patients by Natural Image and Diffusion Model Augmentation
Cobb, Robert (King's College London), Reader, Andrew J. (King's College London)
π― What it does: Utilizing diffusion models and Perlin noise for data augmentation of limited [99mTc] maraciclatide hand images, combined with natural hand images to enhance multi-label classification performance for RA inflammation.
π― What it does: This paper implements cross-domain fetal brain tissue segmentation through a synthetic data generation method based on domain randomization called FetalSynthSeg.
Incorporating Clinical Guidelines through Adapting Multi-modal Large Language Model for Prostate Cancer PI-RADS Scoring
Zhang, Tiantian (Chinese University of Hong Kong), Dou, Qi (Huazhong University of Science and Technology)
CodeClassificationDomain AdaptationKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningImageMultimodalityBiomedical DataMagnetic Resonance Imaging
π― What it does: This paper proposes a general method for integrating the Clinical PI-RADs Guidelines (PICG) into MRI PI-RADs scoring, utilizing a multi-modal large language model (MLLM) for two-stage fine-tuning and injecting PICG information into existing scoring networks through feature distillation to improve scoring accuracy.
π― What it does: A learnable trigger generation method for medical imaging measurement data is proposed, achieving a pre-imaging backdoor attack without compromising reconstruction quality.
CodeClassificationDomain AdaptationExplainability and InterpretabilityConvolutional Neural NetworkTransformerImageBiomedical Data
π― What it does: This paper incorporates the fine-grained knowledge of clinical experts regarding the importance of concepts into the Concept Bottleneck Model (CBM), guiding the model to prioritize clinically relevant concepts during prediction, thereby enhancing the model's interpretability and cross-domain robustness.
Interpretable Representation Learning of Cardiac MRI via Attribute Regularization
Di Folco, Maxime (Helmholtz Munich), Schnabel, Julia A. (Technical University of Munich)
CodeExplainability and InterpretabilityRepresentation LearningAuto EncoderImageBiomedical DataMagnetic Resonance Imaging
π― What it does: A model combining attribute regularization and Soft Introspective Variational Autoencoder (SIVAE) called AR-SIVAE is proposed for interpretable representation learning of cardiac MRI.
π― What it does: A dual student-teacher combination CNN and Transformer semi-supervised image segmentation framework (DSTCT) is proposed for the segmentation of the pubic symphysis and fetal head in obstetric ultrasound images.
CodeSegmentationTransformerPrompt EngineeringImageBiomedical Data
π― What it does: A multi-view weakly supervised tooth segmentation method based on SAM, IOSSAM, is proposed, utilizing 2D bounding box training prompts to achieve 3D IOS tooth segmentation and FDI labeling.
π― What it does: An iterative pseudo-label correction framework based on SAM is proposed to address the domain adaptation problem in unsupervised medical image segmentation.
Iterative Online Image Synthesis via Diffusion Model for Imbalanced Classification
Li, Shuhan (Hong Kong University of Science and Technology), Cheng, Kwang-Ting (Hong Kong University of Science and Technology)
CodeClassificationGenerationData SynthesisConvolutional Neural NetworkDiffusion modelImageBiomedical Data
π― What it does: An iterative online image synthesis framework (IOIS) is proposed to address the class imbalance problem in medical image classification.
π― What it does: This paper proposes an unsupervised brain lesion segmentation method, IterMask 2, which enhances the accuracy of abnormal region detection through iterative spatial mask refinement and frequency domain high-frequency information guided reconstruction.
π― What it does: An interactive framework for joint denoising of electron microscopy images and instance segmentation is proposed, and the feature layers of the two tasks are mutually enhanced through an instance-aware embedding module.
Joint multi-task learning improves weakly-supervised biomarker prediction in computational pathology
El Nahhas, Omar S. M. (TUD Dresden University of Technology), Kather, Jakob Nikolas (EKFZ for Digital Health TU Dresden)
CodeClassificationTransformerImageBiomedical Data
π― What it does: This paper proposes a weakly supervised joint multi-task Transformer for directly predicting two biomarkers, MSI and HRD, from whole slide images, and learns tumor microenvironment features through auxiliary regression tasks.
π― What it does: This study investigates the impact of the composition of pre-training data in self-supervised learning on the performance of downstream tasks in surgical computer vision, and proposes a pre-training strategy based on multi-center and multi-surgery type data.
Knowledge-driven Subspace Fusion and Gradient Coordination for Multi-modal Learning
Zhang, Yupei (University of Hong Kong), Li, Chao (University of Dundee)
CodeClassificationOptimizationSpiking Neural NetworkMultimodalityBiomedical Data
π― What it does: A knowledge-driven subspace fusion and gradient coordination multimodal learning framework is proposed, which splits pathological images and genomic data into tumor subspaces and microenvironment subspaces for joint modeling, aiming to better capture the interaction information between tumors and the microenvironment.
Knowledge-grounded Adaptation Strategy for Vision-language Models: Building a Unique Case-set for Screening Mammograms for Residents Training
Urooj Khan, Aisha (Mayo Clinic), Banerjee, Imon (Arizona State University)
CodeRetrievalDomain AdaptationTransformerSupervised Fine-TuningVision Language ModelContrastive LearningImageTextMagnetic Resonance Imaging
π― What it does: This paper proposes a knowledge-driven selective sampling strategy for adapting visual-language models (VLM) to the medical field (mammography) and enhances image-text retrieval performance through this strategy.
π― What it does: This study proposes a two-stage brain MRI full lifecycle segmentation framework that achieves high-precision segmentation using Knowledge-Guided Prompt Learning (KGPL) based on a pre-trained model;
Language-Enhanced Local-Global Aggregation Network for Multi-Organ Trauma Detection
Yu, Jianxun (Xidian University), Dou, Qi (Huazhong University of Science and Technology)
CodeClassificationObject DetectionTransformerLarge Language ModelPrompt EngineeringImageBiomedical DataComputed Tomography
π― What it does: A language-enhanced local-global aggregation network is proposed for multi-organ trauma detection, integrating local (organ-level) and global features of images, and introducing large language model text embeddings to enhance anatomical and trauma knowledge.
π― What it does: This paper proposes Laplacian Segmentation Networks (LSN), which achieves simultaneous quantification of model (epistemic) and data (aleatoric) uncertainty by performing a Laplace approximation on the posterior of the weights of medical image segmentation models, thereby detecting out-of-distribution (OOD) anomalies in segmentation results.
π― What it does: This study investigates the real-time prediction of multi-focal lung SABR dose distribution using a two-stage latent space framework (LDFormer) based on Transformer and VQ-VAE.
π― What it does: A lightweight boundary-assisted UNet (LB-UNet) is proposed to specifically address the issues of unclear boundaries and resource constraints in skin lesion segmentation.
π― What it does: A DIF-Gaussian framework based on 3D Gaussian distribution is proposed for limited projection (β€10 views) CBCT reconstruction, incorporating Test-Time Optimization (TTO) during inference to enhance generalization capability.
Learning a Clinically-Relevant Concept Bottleneck for Lesion Detection in Breast Ultrasound
Bunnell, Arianna (University of Hawai'i), Shepherd, John A. (University of Hawai'i)
CodeClassificationObject DetectionSegmentationExplainability and InterpretabilityHyperparameter SearchConvolutional Neural NetworkSupervised Fine-TuningImageBiomedical DataUltrasound
π― What it does: A concept bottleneck model based on the BI-RADS vocabulary (BI-RADS-CBM) is proposed, achieving lesion detection, segmentation, interpretable feature prediction, and cancer classification in breast ultrasound images, while supporting concept-level interventions.
Learning from Partial Label Proportions for Whole Slide Image Segmentation
Matsuo, Shinnosuke (Kyushu University), Bise, Ryoma (Kyoto University Hospital)
CodeSegmentationConvolutional Neural NetworkImage
π― What it does: This paper proposes a weakly supervised learning method for tumor subtype segmentation in whole slide images (WSI), utilizing a 'partial label proportion' (LPLP) that only provides the tumor subtype ratio to classify each image patch.
π― What it does: This paper proposes a framework for learning multi-modal partially labeled multi-organ segmentation using cross-modal prior probability graphs;
π― What it does: A model called DeCode has been developed for decoder conditioning using shape features in 3D segmentation tasks, capable of achieving label-free conditioning during inference through learned embeddings.
π― What it does: Using unpaired bone tissue pathological images as a teacher model, knowledge distillation is performed on synchrotron radiation microCT (SR-microCT) images to train a model that only requires SR-microCT for bone cavity segmentation.
π― What it does: A method is proposed to extract task-specific class prototypes based on an image-caption database, and to select WSI annotation areas using prototype similarity, thereby reducing the cost of full-image annotation.
π― What it does: A parameter-efficient fine-tuning framework LGA is proposed, which integrates the text features of medical reports extracted by BERT with the SAM image encoder through a lightweight adapter (Feature Fusion Module) to achieve multi-modal medical image segmentation.
LGRNet: Local-Global Reciprocal Network for Uterine Fibroid Segmentation in Ultrasound Videos
Xu, Huihui (Shanghai Artificial Intelligence Laboratory), Zhu, Lei (Hong Kong University of Science and Technology)
CodeSegmentationVideoBiomedical DataUltrasound
π― What it does: This paper proposes a Local-Global Recursive Network (LGRNet) for the segmentation of uterine fibroids in ultrasound videos, and constructs the UFUV dataset with 100 videos and a total of 5000 annotated frames for the first time.
π― What it does: A lifelong whole slide image retrieval framework (LWSR) is proposed, which achieves continuous learning and avoids catastrophic forgetting through a local memory pool and a Distance Consistency Replay (DCR) module.
π― What it does: A UNet based on large kernel Mamba (LKM-UNet) is proposed for medical image segmentation, utilizing a multi-scale large kernel structure to enhance local and global feature modeling.
LLM-guided Multi-modal Multiple Instance Learning for 5-year Overall Survival Prediction of Lung Cancer
Kim, Kyungwon (Yonsei University), Hwang, Dosik (DoctorWorks Co., Ltd.)
CodeTransformerLarge Language ModelImageTextMultimodalityComputed Tomography
π― What it does: This paper proposes a multimodal multi-instance learning framework guided by LLM, which jointly predicts the 5-year overall survival rate of lung cancer patients using CT, pathological slices, and clinical text information.
π― What it does: This study proposes the LM-UNet framework, which employs a dual-modal (PET+CT) labeling strategy, a CNN-Mamba hybrid encoder, and an anatomical enhancement module to achieve automatic segmentation of whole-body PET-CT lesions, significantly improving DSC and HD95 on both a self-built dual-modal dataset and the public autoPET dataset.
π― What it does: This paper proposes the LEPD-Net model, which combines low-dose abdominal X-rays, position embedding, and pairwise distance learning for fine-grained diagnosis of urinary stones.
π― What it does: A Transformer-based longitudinal medical image analysis framework LOMIA-T was developed to predict the preoperative treatment response (pCR) in patients with esophageal squamous cell carcinoma.
Karaman, Batuhan K. (Cornell University and Cornell Tech), Sabuncu, Mert R. (Cornell University and Cornell Tech)
CodeClassificationTransformerImageBiomedical Data
π― What it does: A Transformer model called LoMaR was constructed and evaluated, capable of receiving an arbitrary number of longitudinal mammography images and predicting future breast cancer risk.
π― What it does: A longitudinally consistent individualized prediction framework is proposed, utilizing surface-based deep networks to generate complete cortical property (thickness, sulcal depth, area, myelination) trajectories of infant scalp at any time point within 0-24 months.
CodeClassificationObject DetectionSegmentationExplainability and InterpretabilityConvolutional Neural NetworkImageBiomedical DataComputed Tomography
π― What it does: A self-supervised framework for diagnosing colorectal cancer with loosely defined lesion locations is proposed, utilizing temporal and cross-modal consistency constraints to achieve precise lesion localization with minimal annotations, and enhancing diagnostic reliability through a masking loop mechanism.
LS+: Informed Label Smoothing for Improving Calibration in Medical Image Classification
Sambyal, Abhishek Singh (Indian Institute of Technology Ropar), Bathula, Deepti R. (Indian Institute of Technology Palakkad)
CodeClassificationConvolutional Neural NetworkImageBiomedical Data
π― What it does: The LS+ (Label Smoothing Plus) method is proposed, which constructs class-specific soft labels using the class accuracy from the validation set to improve the calibration performance of medical image classification.
LSSNet: A Method for Colon Polyp Segmentation Based on Local Feature Supplementation and Shallow Feature Supplementation
Wang, Wei (Changsha University of Science and Technology), Wang, Xin (Chengdu University of Information Technology)
CodeSegmentationTransformerImage
π― What it does: LSSNet is designed for colon polyp segmentation, utilizing two structures: local feature compensation and shallow feature compensation.
π― What it does: A multi-time multi-modal fusion model M2Fusion is proposed, which combines pre/post MRI and whole slide images (WSI) to predict the pathological complete response (pCR) of breast cancer patients after neoadjuvant chemotherapy.
CodeSegmentationTransformerMixture of ExpertsAuto EncoderImageMultimodalityMagnetic Resonance ImagingComputed Tomography
π― What it does: A multi-modal medical image segmentation foundation model M oE based on SwinUNet is proposed, utilizing modality-specific experts and a gating network to achieve cross-modal learning and parameter efficiency.
MAdapter: A Better Interaction between Image and Language for Medical Image Segmentation
Zhang, Xu, Zhang, Lefei (Wuhan University)
CodeSegmentationConvolutional Neural NetworkVision Language ModelImageTextBiomedical DataComputed Tomography
π― What it does: A medical image segmentation framework based on a bidirectional MAdapter is proposed, which utilizes pre-trained visual and language encoders to extract multi-scale features, and achieves the fusion and alignment of visual and textual information through a bidirectional interaction module and a dedicated decoder, ultimately generating pixel-level segmentation results.
MambaMIL: Enhancing Long Sequence Modeling with Sequence Reordering in Computational Pathology
Yang, Shu (Hong Kong University of Science and Technology), Chen, Hao (Harvard Medical School)
CodeClassificationComputational EfficiencyConvolutional Neural NetworkImageBiomedical Data
π― What it does: MambaMIL is proposed within the framework of Multiple Instance Learning (MIL), utilizing Mamba (a state space model with linear time complexity) and the newly designed Sequence Reordering Mamba (SR-Mamba) to model long sequence instances in Whole Slide Images (WSI), thereby enhancing feature extraction and classification/prognosis performance in pathological imaging.
Mammo-CLIP: A Vision Language Foundation Model to Enhance Data Efficiency and Robustness in Mammography
Ghosh, Shantanu (Boston University), Batmanghelich, Kayhan (University of Pittsburgh)
CodeClassificationObject DetectionTransformerVision Language ModelContrastive LearningImageTextBiomedical DataMagnetic Resonance Imaging
π― What it does: We propose Mammo-CLIP, a vision-language foundation model for mammography, and develop Mammo-FActOR for weakly supervised attribute localization and interpretation.
π― What it does: Using the bSSFP MRF sequence, combined with the frequency distribution of simulated microvascular structures, a bidirectional long short-term memory network (BiLSTM) is employed to rapidly reconstruct full-scale quantitative images containing T1, T2, B1, frequency offset, CBV, and vascular radius.
π― What it does: For 3D tumor lesion semantic segmentation, we propose the Mask-Enhanced SAM (M-SAM) framework, which includes a Mask-Enhanced Adapter (MEA) and an iterative refinement scheme; it effectively injects tumor location priors by inserting the adapter into the SAM-Med3D framework and gradually refines the segmentation masks.
Mask-Free Neuron Concept Annotation for Interpreting Neural Networks in Medical Domain
Kim, Hyeon Bae (Kyung Hee University), Kim, Seong Tae (Technical University of Munich)
CodeObject DetectionExplainability and InterpretabilityContrastive LearningImageTextBiomedical DataComputed Tomography
π― What it does: This paper studies a mask-free neuron concept annotation method for medical visual models called MAMMI, which is used to explain the internal decisions of the model.
π― What it does: Based on multi-phase CTA images, CT perfusion (CBF/CBV/Tmax) images are generated using the Masked Residual Diffusion Probabilistic Model (MRDPM).
π― What it does: A Transformer-based medical image classification architecture named Med-Former is proposed, which integrates local and global feature learning with spatial attention fusion;
π― What it does: This paper proposes a general training framework MedContext, which can jointly optimize supervised segmentation tasks and self-supervised context reconstruction tasks without the need for large-scale labeled data or pre-training, thereby improving the performance of 3D medical image segmentation.
π― What it does: This paper proposes a single-source domain generalization method based on random synthesis of frequency domain amplitude spectra for medical image segmentation.
π― What it does: The study proposes the MEGFormer model, which aligns magnetoencephalography (MEG) signals with audio speech to decode perceived speech.
π― What it does: A Cascaded Amortized Latent Diffusion Model (CA-LDM) has been developed to achieve 512Β³ resolution optical coherence tomography (OCT) volume synthesis in memory-constrained environments.
π― What it does: A heart image registration framework called MemWarp based on memory networks is proposed, which can retain local discontinuities of anatomical boundaries without the need for segmentation masks during the inference phase.
π― What it does: This paper proposes MeshBrush, a neural network-based 3D mesh shading method that learns from existing image-to-image (I2I) style transfer models to generate temporally coherent and high-fidelity videos from endoscopic renderings of patient CT scans.
π― What it does: This study investigates a metabolism-aware unsupervised anomaly detection framework, MetaAD, designed to highlight abnormal metabolic regions in 18F-FDG PET images of Parkinson's disease.
π― What it does: Proposed the MetaUNETR model, exploring the effects of various token mixers in 3D multi-organ segmentation, and achieving efficient spatial mixing through the TriCruci layer;
π― What it does: This paper proposes the MiHATP network for super-resolution of pathological images, achieving more accurate cell contour recovery through a dual-branch structure and multi-mixed attention combined with transformation pool contrastive learning.
Misaligned 3D Texture Optimization in MIS Utilizing Generative Framework
Zheng, Jieyu (Hefei University of Technology), Ma, Xiang (Hefei University of Technology)
CodeRestorationGenerationOptimizationTransformerGenerative Adversarial NetworkSimultaneous Localization and MappingOptical FlowImagePoint Cloud
π― What it does: Using camera projection to transfer 3D point cloud textures to 2D texture space, and employing a pre-trained KL-Reg VQ-GAN combined with registration and fusion modules to merge two frames of RGB point clouds, generating high-resolution continuous textures.
Mitigating attribute amplification in counterfactual image generation
Xia, Tian (Imperial College London), Glocker, Ben (Imperial College London)
CodeGenerationData SynthesisFlow-based ModelAuto EncoderImageBiomedical DataElectronic Health Records
π― What it does: This study investigates the use of Deep Structural Causal Models (DSCM) to generate counterfactual images in medical imaging, identifying an attribute amplification issue and proposing a solution called Soft Counterfactual Fine-Tuning (Soft-CFT).
π― What it does: Combining CT images, clinical, blood, and radiomics data, the MMFusion model is proposed to fuse multimodal features using heterogeneous graphs, and to eliminate feature redundancy through a conditional feature-guided diffusion process, achieving the diagnosis of lymph node metastasis in esophageal squamous cell carcinoma.
CodeTransformerVision Language ModelImageMultimodalityBiomedical Data
π― What it does: A multi-question learning (MMQL) framework is proposed for medical visual question answering tasks, which jointly processes multiple questions under the same medical image and utilizes answered questions as prompt information to enhance diagnostic accuracy.
π― What it does: A two-stage deep learning framework called MoCo-Diff is proposed, which first uses a dual-branch Transformer to remove MRI motion artifacts. The resulting motion-corrected images are then used as adaptive priors, employing a control network based on Stable Diffusion for high-quality image restoration, incorporating uncertainty prediction to achieve dynamic weighting, thereby enhancing the reliability and detail preservation of the restoration.
π― What it does: This paper proposes the ModelMix method for cardiac image segmentation using a small amount of scribble annotations. It constructs linear mixed virtual models of different task models and enhances segmentation performance through vicinal risk regularization.
MoME: Mixture of Multimodal Experts for Cancer Survival Prediction
Xiong, Conghao (Chinese University of Hong Kong), King, Irwin (Chinese University of Hong Kong)
CodeClassificationOptimizationExplainability and InterpretabilityConvolutional Neural NetworkTransformerMixture of ExpertsMultimodalityBiomedical Data
π― What it does: For the task of cancer survival prediction, this paper proposes a Biased Progressive Encoding (BPE) framework based on a mixture of multimodal experts, and implements a Mixture of Multimodal Experts (MoME) layer that can alternately fuse pathological Whole Slide Images (WSIs) and genomic features multiple times during the encoding process, dynamically selecting the most suitable expert to enhance prediction performance.
CodeClassificationRecognitionOptimizationTransformerVision Language ModelMultimodalityBiomedical Data
π― What it does: This paper addresses the application of multimodal pre-training models in disease diagnosis and proposes the MoRA (Modal-aware Low-rank Adaptation) method, which solves the missing modality problem and significantly reduces training costs.
π― What it does: A pluggable MoreStyle module is proposed, which enhances the domain generalization performance of single-source medical image segmentation by relaxing low-frequency constraints in the Fourier space, utilizing adversarial style augmentation and uncertainty-weighted loss.