Residual Network based Aggregation Model for Skin Lesion Classification

TL;DR

This paper introduces a novel skin lesion classification method combining residual networks for local feature extraction and Fisher vector encoding for image-level representation, demonstrating effectiveness on a major skin imaging dataset.

Contribution

It proposes a new aggregation algorithm that integrates residual networks with Fisher vector encoding for improved skin lesion classification.

Findings

Effective on ISIC2018 dataset

Improves discrimination in skin lesion diagnosis

Combines deep learning with traditional encoding

Abstract

We recognize that the skin lesion diagnosis is an essential and challenging sub-task in Image classification, in which the Fisher vector (FV) encoding algorithm and deep convolutional neural network (DCNN) are two of the most successful techniques. Since the joint use of FV and DCNN has demonstrated proven success, the joint techniques could have discriminatory power on skin lesion diagnosis as well. To this hypothesis, we propose the aggregation algorithm for skin lesion diagnosis that utilize the residual network to extract the local features and the Fisher vector method to aggregate the local features to image-level representation. We applied our algorithm on the International Skin Imaging Collaboration 2018 (ISIC2018) challenge and only focus on the third task, i.e., the disease classification.