{"title":"多参数肝脏MRI可解释病变分类的相关路径网络","authors":"Fakai Wang , Zhehan Shen , Huimin Lin , Fuhua Yan","doi":"10.1016/j.media.2025.103790","DOIUrl":null,"url":null,"abstract":"<div><div>Liver tumor diagnosis is a key task in abdominal imaging examination, and there are numerous researches on automatic classifying focal liver lesions (FLL). More are in CT and ultrasound and fewer utilize MRI despite unique diagnostic advantages. The obstacles lie in dataset curation, technical complexity, and clinical explainability in liver MRI. In this paper, we propose the Correlation Routing Network (CRN) which takes in 10 MRI sequences and predicts lesion types (HCC, Cholangioma, Metastasis, Hemangioma, FNH, Cyst) as well as imaging features, to achieve both high accuracy and explainability. The CRN model consists of encoding branches, correlation routing/relay modules, and the self-attention module. The independent encoding paradigm facilitates information disentangling, the correlation routing scheme helps redirection and decoupling effectively, and the self-attention enforces global feature sharing and prediction consistency. The model predicts detailed lesion imaging features, promoting explainable classification and clinical accountability. We also identify the signal relations and derive quantitative explainability. Our liver lesion classification model achieves malignant-benign accuracy of 97.2%, six-class accuracy of 88%, and averaged imaging feature accuracy of 84.9%, outperforming popular CNN and transformer-based models. We hope to spark insights for multimodal lesion classification and model explainability.</div></div>","PeriodicalId":18328,"journal":{"name":"Medical image analysis","volume":"107 ","pages":"Article 103790"},"PeriodicalIF":11.8000,"publicationDate":"2025-09-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":"{\"title\":\"Correlation Routing Network for Explainable Lesion Classification in Multi-Parametric Liver MRI\",\"authors\":\"Fakai Wang , Zhehan Shen , Huimin Lin , Fuhua Yan\",\"doi\":\"10.1016/j.media.2025.103790\",\"DOIUrl\":null,\"url\":null,\"abstract\":\"<div><div>Liver tumor diagnosis is a key task in abdominal imaging examination, and there are numerous researches on automatic classifying focal liver lesions (FLL). More are in CT and ultrasound and fewer utilize MRI despite unique diagnostic advantages. The obstacles lie in dataset curation, technical complexity, and clinical explainability in liver MRI. In this paper, we propose the Correlation Routing Network (CRN) which takes in 10 MRI sequences and predicts lesion types (HCC, Cholangioma, Metastasis, Hemangioma, FNH, Cyst) as well as imaging features, to achieve both high accuracy and explainability. The CRN model consists of encoding branches, correlation routing/relay modules, and the self-attention module. The independent encoding paradigm facilitates information disentangling, the correlation routing scheme helps redirection and decoupling effectively, and the self-attention enforces global feature sharing and prediction consistency. The model predicts detailed lesion imaging features, promoting explainable classification and clinical accountability. We also identify the signal relations and derive quantitative explainability. Our liver lesion classification model achieves malignant-benign accuracy of 97.2%, six-class accuracy of 88%, and averaged imaging feature accuracy of 84.9%, outperforming popular CNN and transformer-based models. We hope to spark insights for multimodal lesion classification and model explainability.</div></div>\",\"PeriodicalId\":18328,\"journal\":{\"name\":\"Medical image analysis\",\"volume\":\"107 \",\"pages\":\"Article 103790\"},\"PeriodicalIF\":11.8000,\"publicationDate\":\"2025-09-15\",\"publicationTypes\":\"Journal Article\",\"fieldsOfStudy\":null,\"isOpenAccess\":false,\"openAccessPdf\":\"\",\"citationCount\":\"0\",\"resultStr\":null,\"platform\":\"Semanticscholar\",\"paperid\":null,\"PeriodicalName\":\"Medical image analysis\",\"FirstCategoryId\":\"5\",\"ListUrlMain\":\"https://www.sciencedirect.com/science/article/pii/S1361841525003366\",\"RegionNum\":1,\"RegionCategory\":\"医学\",\"ArticlePicture\":[],\"TitleCN\":null,\"AbstractTextCN\":null,\"PMCID\":null,\"EPubDate\":\"\",\"PubModel\":\"\",\"JCR\":\"Q1\",\"JCRName\":\"COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE\",\"Score\":null,\"Total\":0}","platform":"Semanticscholar","paperid":null,"PeriodicalName":"Medical image analysis","FirstCategoryId":"5","ListUrlMain":"https://www.sciencedirect.com/science/article/pii/S1361841525003366","RegionNum":1,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q1","JCRName":"COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE","Score":null,"Total":0}
Correlation Routing Network for Explainable Lesion Classification in Multi-Parametric Liver MRI
Liver tumor diagnosis is a key task in abdominal imaging examination, and there are numerous researches on automatic classifying focal liver lesions (FLL). More are in CT and ultrasound and fewer utilize MRI despite unique diagnostic advantages. The obstacles lie in dataset curation, technical complexity, and clinical explainability in liver MRI. In this paper, we propose the Correlation Routing Network (CRN) which takes in 10 MRI sequences and predicts lesion types (HCC, Cholangioma, Metastasis, Hemangioma, FNH, Cyst) as well as imaging features, to achieve both high accuracy and explainability. The CRN model consists of encoding branches, correlation routing/relay modules, and the self-attention module. The independent encoding paradigm facilitates information disentangling, the correlation routing scheme helps redirection and decoupling effectively, and the self-attention enforces global feature sharing and prediction consistency. The model predicts detailed lesion imaging features, promoting explainable classification and clinical accountability. We also identify the signal relations and derive quantitative explainability. Our liver lesion classification model achieves malignant-benign accuracy of 97.2%, six-class accuracy of 88%, and averaged imaging feature accuracy of 84.9%, outperforming popular CNN and transformer-based models. We hope to spark insights for multimodal lesion classification and model explainability.
期刊介绍:
Medical Image Analysis serves as a platform for sharing new research findings in the realm of medical and biological image analysis, with a focus on applications of computer vision, virtual reality, and robotics to biomedical imaging challenges. The journal prioritizes the publication of high-quality, original papers contributing to the fundamental science of processing, analyzing, and utilizing medical and biological images. It welcomes approaches utilizing biomedical image datasets across all spatial scales, from molecular/cellular imaging to tissue/organ imaging.