Interpretable AI-Enabled Model for Skin Cancer Diagnosis using LIME

Skin cancer ranks among the most prevalent forms of cancer worldwide. According to statistics reported by the Skin Cancer Foundation, it is estimated that one in every three cancer diagnoses globally pertains to a form of skin cancer. This high incidence underscores the critical need for early detection and effective treatment strategies. In most cases, the cancer is diagnosed in its later stage, thereby raising the mortality rate. If the diagnosis of skin cancer happens in the early stage, some of the lives may be salvaged. A deep learning (DL) based method is proposed for the detection and classification of skin cancer. The whole process is divided into four major stages: data pre-processing, model training, performance evaluation, and result analysis. The model was trained with a dataset publicly available that consisted of 2,357 images in nine categories. The dataset was very small, hence data augmentation with Synthetic Minority Over-Sampling Technique (SMOTE) was applied to expand the dataset. Moreover, since the original dataset did not contain images of non-cancer cases, 98 such images were collected from a local dermatologist to allow the model to classify cancerous and non-cancerous cases. Due to the small dataset, the Convolutional Neural Network (CNN) along with the ResNet50 model were used to perform feature extraction of images. For explaining the model’s predictions, the Local Interpretable Model-Agnostic Explanations (LIME) method was also used. Model evaluation was performed on four configurations: CNN on balanced dataset, CNN on original dataset, ResNet50 on balanced dataset, and a hybrid model on a balanced dataset. From the experimental results, it appears that the CNN on the balanced training database provided the highest performance with 94.5% training accuracy and 90.05% on the test data.