In recent years, many AI models have been developed to aid physicians in the diagnosis of different types of skin cancer. However, little progress has been made in providing accurate diagnoses before meeting a physician, which could potentially save large amounts of time for all parties involved. In this work, we demonstrate the potential of using large model ensembles to provide highly accurate estimations for the presence of skin cancer from a given image. Our best ensemble reached a peak pAUC-above-80 score of 0.171. In addition, we showcase the significant improvement that can be made through various augmenting and preprocessing techniques. Our work also has the novel use of Quadruple Stratified Leak-Free KFold Cross-Validation in medical areas.
Keywords: ai, ml, classification, preprocessing, data augmentation
TL;DR: The paper talks about the the novel use of stratified k-fold validation, data augmentation techniques, and ensemble learning for the accurate classification of skin cancer.
Abstract:
Submission Number: 77
Loading