Skin whole slide image segmentation using lightweight-pruned transformer

Abstract

The American Institute of Dermatology states that among every four individuals, one suffers from skin disease, adding a burden of up to 75 billion dollars in medical health care. Moreover, skin diseases contribute towards the psychological health of the patient making it even more difficult to cater for the disease. Traditional skin diagnosis pipeline initiates with sample extraction using biopsy, followed by chemical staining to enhance disease-associated structures. The stained sample is then further analyzed by the pathologist for diagnosis. To augment the detection of skin whole slide layers (epidermis, dermis, dermo-epidermal junction, keratin/stratum corneum, and slide background), we present a pruned SegFormer architecture (Derma-Pruned). Utilizing self-sufficient attention matrices, Gaussian positional embedding, and adaptive pruning has helped the model learn relevant features and has reduced redundant feature representations. An accuracy of 94.4% has been observed by the updated architecture when trained and tested on 32 whole slide images acquired, stained, and annotated for five layers by pathologists. We have also compared the performance of baseline models trained on unstained, virtually stained, and chemically stained whole slide images. Models trained on stained images performed significantly better than those trained on unstained images, moreover, a high cross-correlation score has been observed on images segmented from models trained using virtually stained and chemically stained images, emphasizing the accuracy of using virtually stained images in the skin disease diagnostics.