A Comparison of Machine Learning Models to Deep Learning Models for Cancer Image Classification and Explainability of Classification

The use of Artificial Intelligence in medical applications is now well established. Deep learning models have proven to be well suited to classification problems in medical applications. Machine learning, and deep learning offer many models to solve such problems. Additionally, as deep learning models become more widely used, the requirement for transparency of decision making, referred to as XAI - or Explainable AI is becoming more urgent for medical applications given both the criticality associated with diagnosis of disease, such as cancer, and the need for transparency of the decision made by the deep learning model so that confidence and trust is increased along with acceptance and more wide deployment of deep learning models in clinical settings. The objective of the current study is to compare the performance of machine learning models to deep learning models, in this case the convolutional Neural network model, when performing a binary classification task for cancer images; cancer or noncancer, then we apply the Grad-CAM algorithm to the trained model as a first step towards explainability of the classification. The investigation was conducted and structured as an experimental study. We used a large dataset from Kaggle, the RSNA-Breast-Cancer-Augmented-Images, owned by Harrison TW White. The results show that machine learning models perform well on small dataset, but not well for large dataset. The XAI experiment shows that the heatmap produced an image that shows a high contrast for the region of interest which may be of help to medical imaging specialists build confidence in the models.