Supervisors: Krishna Chaitanya, Prof. Ender Konukoglu
Whole-body Magnetic resonance images (MRIs) are acquired to infer the different fat and muscle tissues and their composition. These tissue properties and composition are used to evaluate the physiological conditions such as the increased risk of diabetes and cardiovascular diseases by measuring intra-abdominal fat. In recent years, convolutional neural networks (CNNs) have been the popular approach to infer the segmentation mask for MR images in clinical settings. However, CNNs that yield high segmentation accuracy needs a large amount of manually annotated data for training. In practical scenarios, obtaining annotations from experts is time-consuming and expensive. Hence, in this work, we investigate a multi-task learning approach in the limited annotation setting to improve the accuracy over supervised learning of only segmentation task (baseline). For the multi-task learning, we have an additional regressor branch attached to the encoder of an existing CNN, and it takes as input the feature maps of the encoder at various depths obtained from the input image. The regression network is trained for the auxiliary task, where it predicts the relative tissue percentages present in the input 2D slice. Our experiments on a whole-body MRI data yielded small improvements in dice score for multi-task learning approach over the baseline.