A new approach to skin lesion classification with EnTransfer CNN


Soyuslu N. S. , Durusoy Y., Kucukoner M. S.

2021 Innovations in Intelligent Systems and Applications Conference, ASYU 2021, Elazığ, Turkey, 6 - 08 October 2021 identifier

  • Publication Type: Conference Paper / Full Text
  • Doi Number: 10.1109/asyu52992.2021.9599051
  • City: Elazığ
  • Country: Turkey
  • Keywords: cancer, deep learning, ensemble modeling, neural networks, skin lesion, transfer learning
  • Middle East Technical University Affiliated: Yes

Abstract

© 2021 IEEE.Skin cancers are common all over the world. It is expected that an increase will occur in the incidence of skin cancers due to the decrease of the ozone levels in the atmosphere. Early diagnosis of skin cancers increases the survival rate. Many studies have been published that detect skin cancer from dermatoscope images with deep learning algorithms so far. In this study, it is aimed to compare the success of transfer learning models and ensemble transfer learning models in the early diagnosis of skin cancers. 17,571 dermatoscope images containing 9 different skin lesions were used in the study. The data were divided into training, validation, and testing subsets. Then, 7 different transfer learning models were trained and evaluated with these data. Next, the best performing 2-to-7 models were assembled using our proposed ensemble method 'En Transfer' which is based on the stacked ensemble framework. Among the models evaluated with the test data (1760 images), the ensemble model created with the best 5 top-model achieved 93% accuracy (0.9316 weighted mean F1 score), and was the most successful. Among the non-combined models, InceptionResnetV2 achieved 89% accuracy (0.8947 weighted mean F1 score) and was the most successful. The data obtained showed that deep learning models can detect skin cancers with a high success rate. Also, it is shown that our ensemble models' success is much more than a single transfer learning model can achieve. Furthermore, by using the models trained in our study, new applications classifying 9 different lesions can be developed.