Melanoma Skin Cancer Classification via Region-Aware Hierarchical Feature Aggregation

Abstract

Melanoma is a type of skin cancer with the highest risk of death. It is critical to identify these early, as the chances of survival significantly drop after stage 2 melanoma. The use of many technologies has lessened this risk but is still limited when correctly identifying malignant ones. Classifying malignant skin cancers is difficult for the following reasons: the shapes and sizes of melanomas are irregular. It is also challenging to visually distinguish between melanomas and non-melanoma regions. The performance of the previous research is based on the depth of the networks. This has a significant trade-off as using more layers adds to the computational cost of the process. Also, their methods use melanoma segmentation as an auxiliary input to the classification network. Thus, the performance of the classification significantly drops when the segmentation task fails. To address this issue, I propose a novel melanoma classification network that uses hierarchical feature aggregation with an attention mechanism. The overall architecture of the proposed network is as follows: The melanoma feature extractor takes the melanoma image as input and produces the image feature related to melanoma. The second module, the attention network, takes the same melanoma image and outputs the attention map which provides the melanoma feature extractor with feature-level regions of interest. The proposed network achieves an accuracy of 82.7% on the melanoma detection dataset which is publicly available online. Throughout the experiments, I have shown that the proposed method outperforms the previous state-of-the-art methods.