TransUnet for psoriasis lesion segmentation

Outstanding performance of the transformer-based model in the field of natural language processing has piqued the interest of researchers in investigating these techniques for computer vision. And the most popular UNet model is considered a major player in the field of image segmentation. Thus, in this paper, we have proposed the transformer-based UNet model for the complex task of psoriasis lesion segmentation from raw color images. One of the major challenges for our segmentation task is the scarcity of datasets and to overcome this challenge we have exploited the EfficientNetB1 transfer learned model as a backbone for our segmentation model. The proposed model is evaluated for the 70:30 hold-out data division technique and the segmentation performance is evaluated using the Dice Score (DS) and Jaccard Index (JI). The value of DS and JI obtained for the intended task are 0.9571 and 0.9201 respectively with the proposed model. Comparative analysis with different derivatives of the UNet model and state-of-the-art literary work shows the better performance of our proposed model.