Analyzing Dermatological Data for Disease Detection Using Interpretable Deep Learning

Please enable javascript in your browser.

Hashemi Golpaygani, Fatemeh Sadat | 2022

547 Viewed

Type of Document: M.Sc. Thesis
Language: Farsi
Document No: 55012 (19)
University: Sharif University of Technology
Department: Computer Engineering
Advisor(s): Rabiee, Hamid Reza; Sharifi Zarchi, Ali; Ghandi, Narges
Abstract:
We present a deep neural network to classify dermatological disease from patient images. Using self-supervised learning method we have utilized large amount of unlabeled data. We have pre-trained our model on 27000 dermoscopic images gathered from razi hospital, the best dermatological hospital in Iran, along with 33000 images from ISIC 2020 dataset. We have evaluated our model performance in semi-supervised and transfer learning approaches. Our experiments show that using this approach can improve model accuracy and PRC up to 20 percent on semi-supervised setting. The results also show that pretraining can improve classification PRC up to 20 percent on transfer learning task on HAM10000 dataset. We also check the
interpretability of model using GradCAM algorithm. Pre-trained model use some meaningful attribute map to generate feature vector.
Keywords:
Self-Supervised Learning ; Machine Learning ; Interpretability ; Disease Diagnosis ; Dermatological Diseases ; Deep Networks