EXPLAINABLE AI FRAMEWORK FOR MELANOMA DETECTION USING PRE TRAINED CNN FEATURE EXTRACTION AND LOCAL INTERPRETABLE MODEL EXPLANATIONS

Abstract

Melanoma is a deadly skin cancer, the early detection of which can significantly increase its survival, but the manual diagnosis might be subjective and time-consuming. The paper is a proposal to use a combination of feature representations of a trained convolutional neural network (CNN) and local interpretable model explanations (LIME) to categorize dermoscopic images and generate patient-specific pictorial explanations. The research makes use of the benchmark of SIIM-ISIC 2020 Kaggle dataset consisting of 33 126 training images and 10 982 test images of benign and malignant lesions. Two preprocess measures take away artefacts and contrast adjustments and a ResNet-50 network pre-trained on ImageNet extracts deep features. A linear classifier is used to identify melanoma based on these features, and LIME constructs a local surrogate model of each image to suggest parts of the image that are used to reach the decision. Findings indicate that the proposed technique has a precision of 92.7, recall of 94.2, accuracy of 93.4 and AUC of 97.3. We also measure the quality of explanation and confusion matrix values and present the strong and weak points of the framework. This method is a feasible method of integrating high performance classification and interpretable evidence. It is an indicator of directions to come in enhancing fairness and generalization.