DermaTransNet: Where Transformer Attention Meets U-Net for Skin Image Segmentation

Skin, the largest organ of the human body, serves as a protective barrier against the external environment, safeguarding the internal organs. It is composed of layers, namely the epidermis, dermis, and hypodermis. The analysis and detection of diseases heavily rely on identifying the penetration of biomarkers within each layer, thus highlighting the significance of accurate segmentation of each layer in disease diagnosis. In skin disease diagnosis, haematoxylin and eosin-stained whole slide images are the most commonly employed imaging modality. To enhance the efficacy of analysis and assist pathologists, computer-aided diagnostic systems have been proposed to automate the process of disease detection and classification. A crucial initial step in skin disease diagnosis is the accurate segmentation of skin layers, as it facilitates the identification of regions of interest for subsequent processing. This study presents a novel attention-based encoder-decoder architecture comprising a multi-axis encoder with an attention-mixing decoder to precisely segment the skin layers (Epidermis, Dermis, Hypodermis, Keratin) from stained whole slide image samples. The proposed Transformer-based encoder block has helped the model preserve global and local features, which are then passed to the decoder block through an attention-based gated skip connection. The attention mixing decoder module has combined multi-head self-attention, spatial attention, and squeeze excitation modules to enhance the spatial information gain. Additionally, the encoder block aids in extracting high and low-level features, which are further weighted before passing to the decoder block via a gated attention mechanism, thus helping the model acquire the relevant details from all spatial dimensions. The proposed architecture has been evaluated on a publicly available dataset against various evaluation metrics i.e., F1-Score, Accuracy, Precision, Recall, and Hausdorff Distance. The model yielded the highest accuracy of 95.5% when compared with state-of-the-art segmentation models. Additionally, compared with the baseline framework, the model has resulted in an increased overall accuracy of 14.20%.