Efficient-Nets and Their Fuzzy Ensemble: An Approach for Skin Cancer Classification

Abstract

Skin cancer is common and deadly among all cancer types, and its increasing cases in the last decade have put tremendous stress on dermatologists. With the advancement in medical imaging techniques, dermoscopic visual inspection with proper training of dermatologists can achieve approximately 80% diagnostic accuracy. However, in real-life scenarios, most dermatologists ignore the procedural algorithms (3-point checklist, ABCD rule, Menzies method, 7-point checklist) and follow their experience-based instincts. It raises the need for automated dermoscopy diagnosis, and this paper proposes a novel Choquet Fuzzy Ensemble of reward penalized Efficient-Nets for multi-class skin cancer classification. The base classifiers of the architecture are trained with the novel macro F1_score-based rewarding technique to handle the class imbalance of International Skin Imaging Collaboration (ISIC) data. After that, we combine the prediction probabilities of base classifiers using Choquet fuzzy integral to get the final predicted labels. The proposed architecture is evaluated based on ISIC multi-class skin cancer classification. The rewarded cross-entropy loss-based training regime showcased its superiority over weighted cross entropy loss training by attaining 2.61%, 3.06%, and 2.65% improvements in balanced accuracies of base classifiers. The proposed ensemble also outperforms the existing state-of-the-arts in terms of performance. Our model’s highest balanced accuracy (88.15%) over its base classifiers and the state-of-the-art makes our model efficient and trustworthy in the classification goal.

D. Das and N. Arya—Equal contribution.