Publication

Abstract : Skin lesion segmentation is a challenging task in CAD systems due to the variations in shape, size and colour. Fast and accurate skin cancer segmentation can ensure low error rates in clinical assessment of the skin lesions. According to the consensus amongst leading dermatologists, the results produced by the most cutting-edge deep learning segmentation algorithms do not meet the standards for subjective clinical evaluation of skin lesion segmentation. This study introduces a unique, durable, and successful variant of UNet for segmentation of medical images: KP-UNet. The objective is to augment the incorporation of semantic information into fundamental characteristics while simultaneously refining advanced features with greater specificity. This study starts the extraction of several modal features from an input image utilizing a deep neural network encoder. Next, we augment the feature map of each feature set by merging semantic information from each modal feature and these features into a unified representation that progresses from high level to lower-level through the use of the Kronecker product. Our unique skip links endow all tiers' features with rich semantic qualities and fine details. The enhanced characteristics are subsequently sent to the decoder for additional processing and segmentation. The integration of our approach into any Encoder-Decoder network is smooth and seamless. We evaluate the efficacy of our methodology on numerous publically accessible datasets for the segmentation of skin lesions in medical imagery. The experimental findings reveal that our new approach surpasses current contemporary methods regarding segmentation accuracy, while also maintaining memory and computing economy.