Abdominal Adipose Tissue Segmentation in MRI with Double Loss Function Collaborative Learning

Deep learning has shown promising progress in computer-aided medical image diagnosis in recent years, such adipose tissue segmentation. Generally, training a high-performance deep segmentation model requires a large amount of labeled images. However, in clinical practice many labels are saved in numerical forms rather than image forms while relabelling images with manual segmentation is extremely time-consuming and laborious. To fill in this gap between numerical labels and image-based labels, we propose a novel double loss function to train an adipose segmentation model through collaborative learning. Specifically, the double loss function leverages a large volume of numerical labels available and a small volume of images labels. To validate our collaborative learning model, we collect one dataset of 300 high quality MR images with pixel-level segmentation labels and another dataset of 9000 clinical quantitative MR images with numerical labels of the number of pixels in subcutaneous adipose tissue (SAT), visceral adipose tissue (VAT), and Non-adipose tissues. Our approach achieves 94.3% and 90.8% segmentation accuracy for SAT and VAT respectively in the dataset with image labels, and 93.6% and 88.7% segmentation accuracy for the dataset with only numerical labels. The proposed approach can be generalize to a broad range of clinical problems with different types of ground truth labels.