MTA-Net: A Multi-task Assisted Network for Whole-Body Lymphoma Segmentation

Lymphoma treatment planning and prognosis assessment require accurate segmentation of lymphoma lesions. Positron emission tomography (PET) /computed tomography (CT) is widely used for lymphoma segmentation. Many methods do automatic segmentation of lymphoma based on PET/CT. However, a significant challenge that limits the effectiveness of the segmentation method is the large and imbalance variation in size of whole-body lymphoma lesions. For example, a small percentage of images contain large lesions, while most images contain only small lesions or even no lesions, which results in inaccurate segmentation. In this paper, we propose a Multi-task Assisted Network (MTANet) for whole-body lymphoma segmentation. First, we design a novel Multi-task Cross-scale Transformer (MCT) block, which combines the pixels regression task and the whole image classification task at multiple scales. Second, we design a Classification Dynamic Convolution (CDC) whose parameters are additionally controlled by the classification task to assist the segmentation task. In our private whole-body lymphoma dataset, experiments show that MTA-Net achieves the best result among state-of-the-art methods on Dice, HD (Hausdorff Distance), Recall, and Precision.