LAIU-Net: A learning-to-augment incorporated robust U-Net for depressed humans’ tongue segmentation

Computer-aided tongue diagnosis system requires segmentation of the tongue body. The frequent movement of the tongue due to its natural flexibility often causes shape variability in photographs across subjects, which makes segmenting the tongue challenging from non-tongue elements, such as the lips, teeth, and other objects in the background of the tongue. The flexibility of the tongue causes a further challenge in maintaining a similar shape and style when taking photos of many healthy subjects and patients. To address these challenges, we have built a tongue dataset, where the tongue of each subject has been scanned thrice with an interval of less than a second. We have collected 333 tongue images from 111 depressed humans, who have been diagnosed with depression by a psychiatrist. In addition, in this paper, we propose a learning-to-augment incorporated U-Net (LAIU-Net) for the segmentation of the depressed human tongue in photographic images. The best policies for data augmentation were automatically chosen with the proposed LAIU-Net. For this purpose, we corrupted photographic tongue images with the Gaussian, speckle, and Poisson noise. The proposed approach addresses the overfitting problem as well as increases the generalizability of a deep network. We have compared the performance of the proposed LAIU-Net with that of other state-of-the-art U-Net configurations. Our LAIU-Net approach achieved a mean boundary F1 score of 93.1%.