2020/08/05 by Nils Gessert, Julia Krüger, Gessert, Nils +13
Engineering · Immunology and Microbiology · Medicine · #Computer Vision and Pattern Recognition (cs.CV) #FOS: Computer and information sciences #FOS: Electrical engineering #Image Processing Techniques and Applications #Image and Video Processing (eess.IV) #Immunotherapy and Immune Responses #Monoclonal and Polyclonal Antibodies Research #electronic engineering #information engineering
paper · pdf · doi:10.48550/arxiv.2008.02001
openalex publication_date 2020/08/05 · openalex created_date 2022/07/26 · openalex updated_date 2026/07/28
Multiple sclerosis is an inflammatory autoimmune demyelinating disease that\nis characterized by lesions in the central nervous system. Typically, magnetic\nresonance imaging (MRI) is used for tracking disease progression. Automatic\nimage processing methods can be used to segment lesions and derive quantitative\nlesion parameters. So far, methods have focused on lesion segmentation for\nindividual MRI scans. However, for monitoring disease progression,\n\lesion activity in terms of new and enlarging lesions between two time\npoints is a crucial biomarker. For this problem, several classic methods have\nbeen proposed, e.g., using difference volumes. Despite their success for\nsingle-volume lesion segmentation, deep learning approaches are still rare for\nlesion activity segmentation. In this work, convolutional neural networks\n(CNNs) are studied for lesion activity segmentation from two time points. For\nthis task, CNNs are designed and evaluated that combine the information from\ntwo points in different ways. In particular, two-path architectures with\nattention-guided interactions are proposed that enable effective information\nexchange between the two time point's processing paths. It is demonstrated that\ndeep learning-based methods outperform classic approaches and it is shown that\nattention-guided interactions significantly improve performance. Furthermore,\nthe attention modules produce plausible attention maps that have a masking\neffect that suppresses old, irrelevant lesions. A lesion-wise false positive\nrate of 26.4% is achieved at a true positive rate of 74.2%, which is not\nsignificantly different from the interrater performance.\n