International Conference on Medical Image Computing and Computer-Assisted Intervention Β· 609 papers
SPEC-CXR: Advancing Clinical Safety through Entity-Level Performance Evaluation of Chest X-ray Report Generation
Lee, Jung Oh (Lunit Inc.), Kim, Taesoo (Lunit Inc.)
CodeGenerationSafty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBiomedical DataElectronic Health Records
π― What it does: A safety assessment framework SPEC-CXR for generating chest X-ray reports is proposed, which combines large language models for entity-level and report-level structured evaluation of generated reports against reference reports.
π― What it does: This paper proposes a knowledge-enhanced conditional variational autoencoder (KE-CVAE) for directly generating speech waveforms from dynamic MRI sequences, addressing issues of audio noise, distortion, and synchronization in the MRI acquisition environment.
π― What it does: A multi-modal brain disease prediction framework based on multi-relation graph learning and semantic prompting (SPromptGL) is proposed, which constructs a multi-modal relation graph, graph convolutional network, and prompt-guided embedding network to mine the discriminative lesion areas of each modality and fuse global features.
π― What it does: This paper proposes SR-SAM, which utilizes subspace regularization and EMA LoRA for parameter-efficient fine-tuning of the Segment Anything Model to enhance the cross-domain generalization ability of single-source medical image segmentation.
π― What it does: A guided generative model based on implicit neural representations is proposed for synthesizing anatomical structures, particularly the thyroid, which can be controlled for specific patient populations.
π― What it does: Utilizing the CT images from the first time point and the PET images from the second time point for spatiotemporal matching, and generating CT-free attenuation-corrected PET images based on a diffusion model, thereby eliminating the radiation risk associated with the second CT scan in dual-time-point PET/CT scans.
π― What it does: A method for MR bias field correction based on a self-supervised dual network with structural and smoothness constraints (S2DNets) is proposed.
π― What it does: A multimodal MRI missing modality translation framework named MISA-LDM is proposed, which can synthesize the missing MRI modality under any available modality combination and ensures the structural consistency of the generated images through a structure-preserving mechanism.
π― What it does: A Sequential Structure Preserving Extension (SSPE) framework is proposed, which uses a generative network to gradually expand partially overlapping medical images, thereby increasing the overlapping area and improving the registration performance of images with minimal overlap.
Subgroup Performance Analysis in Hidden Stratifications
Bissoto, Alceu (University of Bern), Koch, Lisa M. (University of Bern)
CodeClassificationAnomaly DetectionContrastive LearningImageBiomedical Data
π― What it does: This paper utilizes subgroup discovery techniques for performance analysis in medical image classification, exploring the model performance differences caused by hidden stratification.
π― What it does: A cold-start active learning framework SUGFW based on the Segment Anything Model (SAM) is proposed, which utilizes SAM's zero-shot segmentation capability to estimate uncertainty. It achieves feature weighting and efficient sample selection through Patch-based Global Distinct Representation (PGDR) and Greedy Selection with Cluster and Uncertainty (GSCU) to enhance the labeling efficiency of medical image segmentation.
π― What it does: This paper presents Surface Vision Mamba (SiM), an efficient visual model for spherical surface data aimed at predicting neonatal brain development phenotypes.
π― What it does: A multi-axis longitudinal decoupling framework based on cortical surface maps is proposed, utilizing self-supervised contrastive learning to separate aging effects from multiple Alzheimer's disease trajectories, forming an interpretable latent space.
Surgical Action Planning with Large Language Models
Xu, Mengya (Chinese University of Hong Kong), Dou, Qi (Huazhong University of Science and Technology)
CodeTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelVideoTextMultimodality
π― What it does: This paper proposes a surgical action planning (SAP) task for cholecystectomy and constructs the CholecT50-SAP dataset, utilizing large language models to generate future surgical actions and interpretable text outputs.
SurgX: Neuron-Concept Association for Explainable Surgical Phase Recognition
Kim, Ka Young (Kyung Hee University), Kim, Seong Tae (Technical University of Munich)
CodeRecognitionExplainability and InterpretabilityTransformerVision Language ModelVideo
π― What it does: Proposes the SurgX framework, which provides conceptual explanations for the neurons in the surgical video stage recognition model, constructs a dedicated concept set, and enhances explanation quality through representative sequences.
π― What it does: A weakly supervised semantic segmentation model based on synchronous suppression and activation (SIA-WSSS) is proposed for pathological images.
π― What it does: This paper proposes a collaborative guidance-based regional supervision framework SGRS-Net for semi-supervised medical image segmentation, enhancing the utilization of pseudo-labels.
π― What it does: A 3.5D conditional diffusion model is proposed to synthesize dual-channel color Doppler cardiac ultrasound and B-mode images, and synthetic data is used to enhance RHD detection performance.
π― What it does: An iterative deep network, iterU-Net, is proposed to synthesize multi-time point delayed phase-enhanced images from early dynamic contrast-enhanced breast MRI images, achieving dynamic enhancement curve prediction.
π― What it does: Using synthetic brain MRI ground truth counterfactual data generated by SimBA to evaluate and debug the effects of causal generative models in medical imaging.
π― What it does: The first publicly available multi-class dataset for the risk of muscle invasion in bladder cancer (T2WI-BCMIC) has been proposed and released, and a benchmark model has been constructed based on this dataset. Subsequently, a constraint optimization-based search data augmentation algorithm has been proposed to enhance image classification performance.
π― What it does: This paper proposes the DynamicDPS framework, which combines unconditional diffusion models with data consistency. It first uses a conditional model to generate high-field predictions of low-field MRI, and then suppresses artifacts through dynamic diffusion posterior sampling, achieving high-quality reconstruction of low-field MRI.
π― What it does: A detector-guided adversarial diffusion attack framework (DADA) is proposed, which trains a diffusion model that generates only backgrounds and injects adversarial perturbations during the diffusion process to synthesize high-value false positive samples to enhance polyp detection accuracy.
π― What it does: This paper proposes a low-rank parameter adaptation method based on tensor CUR decomposition, called tCURLoRA, for efficient fine-tuning in medical image segmentation.
π― What it does: A dynamic adaptation framework TEGDA based on test-time evaluation is proposed for test-time adaptation in medical image segmentation across different domains.
π― What it does: A deep learning framework based on a temporal atlas is proposed to generate individualized longitudinal (cross-age) anatomical morphology sequences from simple cross-sectional CT data.
π― What it does: A diffusion framework Mo-Diff based on image-to-video (I2V) is proposed, which predicts future four-dimensional medical image sequences from a single frame.
Temporal Model-Based Federated Active Medical Image Classification
Yan, Yunlu (Hong Kong University of Science and Technology (Guangzhou)), Zhu, Lei (Hong Kong University of Science and Technology)
CodeClassificationFederated LearningImageBiomedical Data
π― What it does: A federated active learning framework for multi-institutional medical image classification, TM-FAL, is proposed, which utilizes temporal local and global model estimates of sample uncertainty and mitigates class imbalance through pseudo-label grouping.
π― What it does: A time-based neural cellular automata (TeNCA) model is used to model dynamic contrast enhancement in breast MRI and achieve synthetic prediction for time-sparse, non-uniformly sampled sequences.
π― What it does: This paper proposes a multi-task self-supervised pre-training framework that learns representations from the temporal trajectories of breast MRI follow-up sequences and predicts whether pathological complete response (pCR) is achieved after axillary neoadjuvant chemotherapy.
TemSAM: Temporal-aware Segment Anything Model for Cerebrovascular Segmentation in Digital Subtraction Angiography Sequences
Zhang, Liang (Huazhong University of Science and Technology), Yang, Xin (Shenzhen University)
CodeSegmentationImageVideoBiomedical Data
π― What it does: This paper proposes TemSAM, a time-aware segmentation model for digital subtraction angiography sequences, which significantly improves vascular segmentation accuracy by combining MIP global cues and complementary frame fusion.
π― What it does: A two-stage network called TESLA is proposed, which achieves multi-contrast brain MRI through-plane super-resolution using progressive reconstruction and structural enhancement, and does not require reference images during inference.
π― What it does: A test-time training method for domain generalization in retinal vessel segmentation is proposed, which generates synthetic images through local contrast-preserving copy-paste (L2CP) during testing.
π― What it does: The MTamba multi-task network is proposed to achieve three-dimensional segmentation of brain gliomas, IDH gene typing, and joint prediction of grading.
π― What it does: A Text-SemiSeg framework is proposed, which combines text information with 3D medical image segmentation to achieve semi-supervised learning.
Text-Guided Multi-Instance Learning for Scoliosis Screening via Gait Video Analysis
Li, Haiqing (University of Texas at Arlington), Huang, Junzhou (University of Texas at Arlington)
CodeClassificationRecognitionTransformerLarge Language ModelVideoText
π― What it does: This paper proposes a text-guided multi-instance learning network (TG-MILNet) for detecting early scoliosis in adolescents through gait videos.
CodeSegmentationTransformerLarge Language ModelImageTextMultimodalityBiomedical DataMagnetic Resonance Imaging
π― What it does: This study created the first volumetric text-image brain tumor segmentation dataset, TextBraTS, and proposed a 3D brain tumor segmentation framework that utilizes sequential cross-attention to integrate text information based on this dataset.
π― What it does: The system evaluates the impact of large-scale pre-training on 3D medical object detection, comparing supervised and various self-supervised methods, and verifies their enhancement effects on different detection architectures.
π― What it does: This paper proposes a Residual-Posterior Line Graph Network (RP-LGN) that transforms brain functional connections (edges) into nodes for learning, enhancing the diagnostic capability for neuropsychiatric disorders such as autism and ADHD.
π― What it does: The AutoSAME framework is proposed to achieve automated measurement of left ventricular indicators in cardiac ultrasound based on SAM.
This EEG Looks Like These EEGs: Interpretable Interictal Epileptiform Discharge Detection With ProtoEEG-kNN
Tang, Dennis (Duke University), Rudin, Cynthia (Harvard Medical School)
CodeClassificationExplainability and InterpretabilityConvolutional Neural NetworkTime SeriesBiomedical Data
π― What it does: This paper proposes the ProtoEEG-kNN model for interpretable detection of interictal epileptiform discharges (IEDs) and provides diagnostic explanations through case comparisons.
tHPM-LDM: Integrating Individual Historical Record with Population Memory in Latent Diffusion-based Glaucoma Forecasting
Fan, Yuheng (University of Liverpool), Zhao, He (Liverpool University Hospitals Nhs Foundation Trust)
CodeClassificationGenerationTransformerDiffusion modelContrastive LearningImageTime SeriesBiomedical Data
π― What it does: A glaucoma prediction framework based on a conditional latent diffusion model (t HPMβLDM) is proposed, which captures the evolution of individual irregular follow-up records using continuous-time attention and introduces population memory modules to incorporate group progression information, enabling predictions of future retinal images and disease classification.
π― What it does: Proposes the Tied Prototype Model (TPM), which improves ADNet for few-shot medical image segmentation, supports multiple prototypes and multiple categories, and introduces ideal threshold estimation;
Time-Lapse Video-Based Embryo Grading via Complementary Spatial-Temporal Pattern Mining
Sun, Yong (Hong Kong University of Science and Technology (Guangzhou)), Nie, Qiang (Hong Kong University of Science and Technology (Guangzhou))
CodeClassificationRecognitionTransformerMixture of ExpertsVideo
π― What it does: A time-lapse video-based embryo grading framework called CoSTeM has been designed and implemented to automatically predict the overall quality grading of artificially fertilized embryos.
TMSE: Tri-Modal Survival Estimation with Context-aware Tissue Prototype and Attention-Entropy Interaction
Zhang, Ruofan (University of Chinese Academy of Sciences), Dong, Di (University of Chinese Academy of Sciences)
CodeClassificationData-Centric LearningLarge Language ModelMultimodalityBiomedical Data
π― What it does: A tri-modal survival prediction framework TMSE is proposed, integrating whole slide images (WSI), pathology reports, and genomic data. This is achieved through specialized report preprocessing, a Context-aware Tissue Prototype (CTP) module, and an Attention-Entropy Interaction (AEI) module to predict patient prognosis time.
Top-Down Attention-based Multiple Instance Learning for Whole Slide Image Analysis
ReisenbΓΌchler, Daniel (University of Regensburg), Merhof, Dorit (Hannover Medical School)
CodeClassificationTransformerImage
π― What it does: A two-stage multi-instance learning framework TDA-MIL is proposed, combining self-attention and task-specific feature selection to achieve efficient classification of Whole Slide Images.
π― What it does: This paper proposes a lightweight topological constraint learning framework called TopoNet, aimed at achieving high precision and efficiency in landmark detection during laparoscopic liver surgery.
Towards Accurate Tumor Budding Detection: A Benchmark Dataset and A Detection Approach Based on Implicit Annotation Standardization and Positive-Negative Feature Coupling
Sun, Rui-Qing (Beijing Institute Of Technology), Zhang, Shaohui (Beijing Institute Of Technology)
π― What it does: Developed a Tumor Budding Detection Network (TBDNet) based on YOLOv5 and released the first publicly available tumor budding detection benchmark dataset, TBDD.
π― What it does: This paper presents a new surgical scene graph dataset, EndoscapesβSG201, and proposes a novel scene graph generation model, SSGβCom, which can simultaneously capture the tool-action-target triplet relationships and hand identity information to enhance the understanding of surgical scenes.
π― What it does: Using a 2D multimodal diffusion model, 3D CT is treated as an ordered sequence of 2D slices to achieve 3D CT synthesis with variable axial lengths.
Training state-of-the-art pathology foundation models with orders of magnitude less data
Karasikov, Mikhail (kaiko.ai), OtΓ‘lora, Sebastian (Netherlands Cancer Institute)
CodeClassificationSegmentationRepresentation LearningTransformerContrastive LearningImageBiomedical Data
π― What it does: Trained a small number (12kβ92k) of WSI pathology visual foundation models, demonstrating that high performance can be achieved with less data.
π― What it does: A Transformer-based LACT sinogram domain segmentation model called TransSino is proposed, which can achieve high-quality lesion segmentation in the absence of missing projection angles.
Treat: A Unified Text-guided Conditioned Deep Learning Model for Generalized Radiotherapy Treatment Planning
Kim, Sangwook (University Health Network), McIntosh, Chris (University Health Network)
CodeTransformerBiomedical Data
π― What it does: A unified text-guided deep learning model, Treat, is proposed to predict dose distribution under various radiotherapy protocols, supporting personalized automatic radiotherapy planning.
π― What it does: Two types of semantic loss functions based on label trees are proposed for sparse annotation multi-class hyperspectral image segmentation, and they are integrated with an OOD detection framework.
CodeClassificationDomain AdaptationTransformerVision Language ModelContrastive LearningBiomedical DataMagnetic Resonance Imaging
π― What it does: This paper introduces Segmented Conformal Prediction (SCP) into few-shot medical visual language models and proposes Unsupervised Transductive Segmented Conformal Adaptation (SCA-T) to maintain coverage guarantees and improve the efficiency of the prediction set.
π― What it does: A fine-tuning framework based on the tumor microenvironment (TME) for a pathological foundation model (PFM) has been designed and implemented to predict the response of esophageal squamous cell carcinoma (ESCC) patients to immunotherapy.
π― What it does: This paper proposes a method for extracting tumor heterogeneity using DCE-MRI time intensity curve clustering, and achieves tumor segmentation on non-contrast breast MRI through a predictive model and a fusion segmentation network.
π― What it does: A two-stage conditional generative model, AneuG, was designed and implemented to synthesize brain aneurysm meshes and their parent vessels that meet specific morphological parameters.
π― What it does: A task vector-based 3D medical segmentation model merging framework is proposed, and the impact of pre-training strategies on merging performance is analyzed in depth.
U-RWKV: Lightweight medical image segmentation with direction-adaptive RWKV
Ye, Hongbo (University of Science and Technology of China), Zhou, S. Kevin (Hohai University)
CodeSegmentationConvolutional Neural NetworkImageBiomedical Data
π― What it does: Proposes the U-RWKV lightweight medical image segmentation framework, utilizing the RWKV structure and adaptive modules to achieve efficient long-range dependency modeling.
CodeConvolutional Neural NetworkBiomedical DataMagnetic Resonance ImagingFibre Orientation Distribution
π― What it does: This paper presents UFO-3, an unsupervised three-compartment model combined with a U-Net framework for estimating the fiber orientation distribution function (fODF), which can achieve high-precision fODF estimation using only single-shell data.
π― What it does: This paper proposes an unsupervised framework that combines implicit neural representations (INR) with neural style transfer (NST) to transfer structural details from high-field MRI to ultra-low-field MRI, enhancing image quality.
UltraRay: Introducing Full-Path Ray Tracing in Physics-Based Ultrasound Simulation
Duelmer, Felix (Technical University of Munich), Navab, Nassir (Technische UniversitΓ€t MΓΌnchen)
CodeMeshComputed TomographyUltrasound
π― What it does: A full-path ray tracing-based ultrasound physical simulation framework, UltraRay, is proposed, which can generate realistic B-mode images in real-time.
π― What it does: Using sparse multi-view 2D ultrasound images, a 3D cardiac anatomical digital twin is constructed through a diffusion transformer and an implicit autoencoder.
Wysocki, Magdalena (Technical University of Munich), Azampour, Mohammad Farid (Technical University of Munich)
CodeNeural Radiance FieldImageMeshUltrasound
π― What it does: Utilizing the acoustic features of multi-view ultrasound B-mode images, a occupancy network is constructed to achieve 3D anatomical structure reconstruction.
CodeDomain AdaptationKnowledge DistillationConvolutional Neural NetworkImageBiomedical Data
π― What it does: A framework called 'Uncertainty-Aware Multi-Expert Knowledge Distillation (UMKD)' is proposed to transfer knowledge from multiple expert models to a single student model, addressing the issues of class imbalance and domain transfer in medical image classification.
π― What it does: This paper proposes a tri-modal MRI fusion framework based on uncertainty perception (UMMF) for predicting HIV-related asymptomatic neurocognitive impairment (ANI).
CodeSegmentationExplainability and InterpretabilityImageBiomedical DataMagnetic Resonance Imaging
π― What it does: By incorporating self-supervised uncertainty loss based on image gradients and noise into the evidence deep learning segmentation model, the interpretability and robustness of uncertainty are enhanced.
π― What it does: Designed and evaluated a balanced multimodal learning framework called UniCross for Alzheimer's disease diagnosis and MCI conversion prediction.
UniOCTSeg: Towards Universal OCT Retinal Layer Segmentation via Hierarchical Prompting and Progressive Consistency Learning
Zhong, Jian, Tang, Xiaoying (Chinese University of Hong Kong)
CodeSegmentationConvolutional Neural NetworkTransformerPrompt EngineeringBiomedical Data
π― What it does: This paper proposes a universal OCT retinal layer segmentation framework called UniOCTSeg, which can uniformly perform segmentation tasks on multi-granularity labeled datasets.
π― What it does: A unified lesion segmentation framework called UniSegDiff is proposed, which employs a staged diffusion model for training and inference.
π― What it does: A pipeline for brainstem shape processing based on Statistical Shape Analysis (SSA) was developed to detect Joubert syndrome using MRI surface data.
π― What it does: An unsupervised cardiac video translation model MFD-V2V has been developed, capable of converting low signal-to-noise ratio and low-contrast DENSE CMR sequences into high-quality, high-contrast cine CMR sequences.
Unsupervised Discovery of Spatiotypes and Context-Aware Graph Neural Networks for Modeling Clinical Endpoints
Dawood, Muhammad (University of Oxford), Rittscher, Jens (University of Oxford)
CodeGraph Neural NetworkBiomedical Data
π― What it does: This paper proposes a Band Descriptor based on concentric rings to quantify the relative abundance of different cell types surrounding single cells. Utilizing this descriptor, substructures of the local cellular microenvironment (spatiotypes) were discovered in spatial transcriptomic samples of pulmonary fibrosis, and these were incorporated as node features into a graph neural network (GNN) to enhance the accuracy of clinical outcome predictions.
Unsupervised Quality Control and Enhancement of Polyp Segmentation in Colonoscopy Videos using Spatiotemporal Consistency
Li, Yujia (Nanjing University of Science and Technology), Zhang, Yizhe (Nanjing University of Science and Technology)
CodeSegmentationVideo
π― What it does: This paper proposes an unsupervised quality control and enhancement framework for polyp segmentation in colonoscopy videos, utilizing the video segmentation base model SAM2 to propagate segmentation results over time, calculate the spatiotemporal consistency score (SQA), and perform re-segmentation of low-quality frames based on high-quality frames.
CodeObject DetectionExplainability and InterpretabilityTransformerLarge Language ModelVision Language ModelImageTextMultimodalityBiomedical DataChain-of-Thought
π― What it does: The V2T-CoT method is proposed, which applies Vision-to-Text Chain-of-Thought (CoT) to medical visual question answering (Med-VQA) by automatically locating disease-related areas and generating interpretable diagnostic reasoning paths based on this.
VAMPIRE: Uncovering Vessel Directional and Morphological Information from OCTA Images for Cardiovascular Disease Risk Factor Prediction
Wang, Lehan (Hong Kong University of Science and Technology), Li, Xiaomeng (Southern Medical University)
CodeClassificationSegmentationConvolutional Neural NetworkTransformerLarge Language ModelImageBiomedical Data
π― What it does: A multi-task framework called VAMPIRE based on OCTA images is proposed, which can simultaneously predict the 10-year risk of cardiovascular disease and four blood indicators (high blood sugar, high cholesterol, high triglycerides, high blood pressure);
VAP-Diffusion: Enriching Descriptions with MLLMs for Enhanced Medical Image Generation
Huang, Peng (Fudan University), Guo, Yi (Fudan University)
CodeGenerationData SynthesisLarge Language ModelPrompt EngineeringDiffusion modelImageBiomedical DataChain-of-Thought
π― What it does: This paper proposes a framework for generating medical images by utilizing multimodal large language models to generate visual attribute prompts, combined with diffusion models;
π― What it does: In medical image classification and segmentation tasks, a method is proposed to introduce Variational Visible Layers (VVL) at the first and/or last layer of the network, which allows for post-hoc fine-tuning on existing models or joint training, achieving lightweight uncertainty estimation.
Various Attention Mechanism Graph Convolutional Network with Multi-Source Domain Adaptation for Cross-Subject EEG Emotion Recognition
Shi, Shuo (Hebei University of Technology), Hao, Xiaoke (Hebei University of Technology)
CodeRecognitionDomain AdaptationGraph Neural NetworkBiomedical Data
π― What it does: This paper proposes a model that combines various attention mechanisms with graph convolutional networks and multi-source domain adaptation for cross-subject emotion recognition.
π― What it does: This paper proposes a voxel-based coarse-to-fine two-stage framework (VBCD) that can automatically generate personalized dental crowns from intraoral scanning data.
π― What it does: A dual-modal segmentation and active learning framework based on vector quantization (VQ-BEGAL) is proposed, achieving feature decoupling and reliable uncertainty estimation.
π― What it does: A framework for autoregressive generation of vascular geometry, called VesselGPT, is proposed. It first learns a discrete vocabulary of node attributes using VQ-VAE, then generates a sequence of nodes using GPT-2, and finally decodes the sequence into a three-dimensional vascular network.
π― What it does: Proposes the Vision-Amplified Semantic Entropy (VASE) method for detecting model hallucination answers in medical visual question answering (VQA).
π― What it does: This paper proposes a lightweight dual-path network called VisNet, which draws inspiration from the human visual system's radiating cell pathways, primary visual cortex, and dorsal/ventral pathways. It designs dual-path multi-scale perception, edge detection and shape adaptation, and spatial semantic extraction modules to achieve efficient denoising of CT, X-ray, and MRI images.
π― What it does: A segmentation-free Vision Transformer-based ViTAL-CT framework is proposed to classify high-risk plaques directly on the CTA plaque cross-sections aligned with the coronary artery centerline.
ViTexNet: Vision-Text Guided Dynamic Convolution Network for Medical Image Segmentation
Bhardwaj, Rahul (Indian Institute of Technology Guwahati), Neog, Debanga Raj (Indian Institute of Technology Guwahati)
CodeSegmentationConvolutional Neural NetworkVision Language ModelImageTextMultimodalityBiomedical DataComputed Tomography
π― What it does: A multimodal medical image segmentation network ViTexNet based on visual-text dynamic convolution is proposed, utilizing text information to guide convolution for efficient segmentation.
VoxelOpt: Voxel-Adaptive Message Passing for Discrete Optimization in Deformable Abdominal CT Registration
Zhang, Hang (Cornell University), Liu, Min (Hunan University)
CodeOptimizationImageComputed Tomography
π― What it does: This paper presents VoxelOpt, a discrete optimization framework based on voxel adaptive information transfer for elastic registration of abdominal CT images.
π― What it does: A self-supervised low-dose CT denoising framework VQ-SCD is proposed, which utilizes positive-dose CT training to achieve low-dose image denoising under unknown scanning conditions.
π― What it does: A variable time-step pulse neural network (VT-SNN) is proposed, which can dynamically terminate inference based on sample uncertainty, improving the efficiency and accuracy of brain magnetic resonance imaging diagnosis.
π― What it does: A new metric named WASABI is proposed and validated to assess the morphological reliability of generated brain MR images, and it is used to compare the performance of various advanced generative models.
π― What it does: WaveFormer is proposed, a 3D Transformer that utilizes Discrete Wavelet Transform (DWT) to perform self-attention at low frequencies and uses Inverse Discrete Wavelet Transform (IDWT) to recover details at high frequencies, aimed at medical image segmentation.
π― What it does: An end-to-end accelerated multi-parameter MRI reconstruction and parameter mapping network based on wavelet separation and physical constraints (WDPM-Net) is proposed.
π― What it does: A hierarchical network WDNet based on wavelet decomposition and diffusion reconstruction is proposed for precise segmentation of polyps and surgical instruments in colonoscopy images.