Triplet-Uncertainty in Multi-Task Deep Learning for Improving Malignancy Characterization of Hepatocellular Carcinoma.

Tumor segmentation and classification have important value in clinical diagnosis. Multi-task learning has been widely used for jointly tumor segmentation and classification in broad medical image scenarios. However, due to the presence of noise in medical image, the performance of multi-task learning model will be inevitably affected. Current research does not consider the impact of heteroscedastic(data) uncertainty and model uncertainty on multi-task model simultaneously. In this work, we propose a triplet-uncertainty in multi-task deep learning network (TU-MTL) for tumor segmentation and classification. Specifically, in addition to the uncertainty estimation of the subtask weight coefficient in multi-task learning, we also consider the heteroscedastic uncertainty estimation and model uncertainty estimation. Furthermore, we consider assigning the weight coefficient of the subtask of uncertainty estimation to the corresponding data uncertainty and model uncertainty of each task. The segmentation and malignancy classification of clinical hepatocellular carcinoma (HCC) show the effectiveness and advantages of the proposed method, especially the introduction of heteroscedastic uncertainty greatly improving the performance.