Reducing annotation effort in muscle MRI segmentation through cross-anatomy learning

Deep learning (DL) has become a cornerstone approach for medical image segmentation, offering high accuracy in various applications. Supervised learning, which relies on labeled datasets, is commonly used for training of DL-based segmentation models targeting a particular prespecified application or a set of such applications. However, a major challenge in clinical deployment is the limited generalizability of these models across varying imaging conditions and anatomical contexts not included or not properly represented in the training set. Transfer learning (TL) tries to overcome this by leveraging knowledge from a source task to improve performance on a related target task within the same or a different image domain. Most TL research focuses on domain adaptation involving differences in imaging protocols or imaging modalities. Less attention has been given to task transfer involving cross-anatomy shifts, where models trained on one anatomical region are applied to a structurally distinct region. This is particularly challenging, as it requires models to adapt to unfamiliar anatomical features. Task transfer is gaining attention for its potential in segmenting different anatomical structures and several studies support the effectiveness of TL for this purpose, especially when target data is limited. In contrast to TL, multi-task learning (MTL) enables simultaneous optimization of related tasks, potentially improving generalization by sharing representations. In this study, we investigate whether knowledge gained from a DL-based model for segmentation of 18 muscles in the proximal leg can be transferred and leveraged to segment 10 muscles in the distal leg within the same image domain, in our case Dixon MRI images acquired with the same scanner and imaging protocol. Although both regions contain long muscles surrounding one or two bones with similar tissue intensities, they involve different anatomical structures and labels. Our goal is to reduce the manual effort required to initiate DL-based segmentation in a new anatomical region. We evaluate supervised TL strategies for transferring knowledge from proximal to distal leg segmentation and explore the potential of MTL to further enhance performance.