Evaluation of State-of-the-Art Concept-Based Explainable Artificial Intelligence Methods in Extraction of Meaningful Information from Oral Cancer Data

Sammanfattning

Though deep neural networks show great potential in many areas, they are often held back due to their inherent lack of transparency. Methods within the field of explainable artificial intelligence have been developed to tackle this problem. Saliency methods among them attempt to point to the areas in the input images that are important for decisions. More recently so called concept-based methods have been developed. These aim to go beyond where in the image important features are located, to give a deeper understanding of what the important features are. In this work we have selected two state-of-the-art concept-based methods with the goal to evaluate them in terms of their ability to extract meaningful information from a deep neural network trained on oral cancer data. We take a functionally-grounded evaluation approach using a synthetic data analogue for real-world data. We put the methods through a model parameter randomization test and a local precision estimation test. We find that both methods appear to find the features we intend. However, there remains challenges in inferring the precise meaning of each method's explanations with great certainty.