A novel explainable image classification framework: case study on skin cancer and plant disease prediction

An explainable/interpretable machine learning model is able to make reasoning about its predictions in understandable terms to humans. These properties are essential in order to trust model’s predictions, especially when these decisions affect critical aspects such as health, rights, security, and educational issues. Image classification is an area in machine learning and computer vision in which convolutional neural networks have flourished since they have shown remarkable performance in such problems. However, these models suffer in terms of transparency, interpretability and explainability, considered as black box models. This work proposes a novel explainable image classification framework applying it on skin cancer and plant diseases prediction problems. This framework combines segmentation and clustering techniques aiming to extract texture features from various subregions of the input image. Then, a feature filtering and cleaning procedure is applied on these extracted features in order to ensure that the proposed model will be also reliable and trustful, while these final extracted features are utilized for training an intrinsic linear white box prediction model. Finally, a hierarchy-based tree approach was created, in order to provide a meaningful interpretation of the model’s decision behavior. The experimental results have shown that the model’s explanations are clearly understandable, reliable, and trustful. Furthermore, regarding the prediction accuracy, the proposed model manages to achieve almost equal performance score (1–2% difference on average) comparing to the state-of-the-art black box convolutional image classification models. Such performance is considered noticeably good since the proposed classifier is an explainable intrinsic white box model.