Multiscale Feature Fusion Using Hybrid Loss for Skin Lesion Segmentation

Skin cancer is prevalent these days but patients’ chances of survival are significantly increased by its early diagnosis. Skin lesion segmentation is an important step in distinguishing skin cancer from malignant lesions in dermoscopic images. Despite the significant advancements made by deep models in enhancing the accuracy and resilience of image segmentation, the task of achieving segmentation outcomes with precise boundaries and accurate structures remains a formidable challenge. In this paper, we proposed a method that combines a CNN architecture and hybrid loss to achieve image segmentation results of high accuracy. The proposed methodology aims to enhance the lesion boundary and integrates low-level details and advanced levels of semantics through the utilisation of feature maps at a range of scales. It also learns hierarchical representations. The model also acquires hierarchical representations by utilising the entire aggregated feature maps. Here, The features of skin lesions are extracted from the segmented region in order to evaluate the lesion. To evaluate the various level of quality of the lesion segmentation, we utilised commonly used evaluation metrics such as the Jaccard Index, IOU, Precision, Dice Coefficient(F1-score), Recall, and Accuracy. Based on the results of the lesion segmentations experiment on ISIC-2018 dataset, we did a comparison of the above metrics with other state-of-the-art techniques. Our proposed method demonstrates exceptional performance across all metrics and achieved better results.