A Robust Explainable AI (XAI) Technique for Tumour Classification

点击翻译按钮获取中文摘要

Recent developments in deep learning have sparked an interest in more high-stakes applications such as medical diagnostics. Given a medical scan, a clinician may want to differentiate between healthy and unhealthy tissue, or between pathologies. However, the black-box nature of DL models means their conclusions tend not to be trusted by clinicians who cannot determine how the model came to its decision. Explainable AI (XAI) techniques exist to bridge this gap by intuitively highlighting the most important features of an input. This gives the end-user an idea of how the model came to its conclusion. Knowing that a model's conclusion is correct is essential in medical diagnostics as their outcomes could impact lives.This project is based on expanding my UG4 Dissertation. During that project, I applied 3 XAI techniques to a medical imaging domain, in order to assess their reliability and usability in the diagnostic field, specifically breast tumour classification. I compared explanations generated by the techniques to each other, and to the opinions of 2 radiologists, and discovered that the techniques disagreed both with each other and with the medical truth. The techniques highlighted conflicting regions as most important to the classification decision. They also highlighted fragmented tumour regions, along with many irrelevant regions, and were deemed unhelpful for diagnostics by our clinicians. This work was also published in the 2022 iMIMIC workshop due to its relevance to the active XAI community.During this PhD I plan to further investigate the shortcomings of XAI techniques in the medical field, and generate a more robust piece of work which showcases a larger number of techniques and ML models, with a much more detailed clinician involvement. I then plan to use my findings to create an XAI technique which performs more reliably. Most techniques involve image segmentation - I plan to experiment with finding segmentation techniques that highlight clinical features, rather than just areas of an image with a similar appearance. I also plan to expand this research to other data modalities, for example gene sequences and tabular patient data.The overall goal of this project is to bring to light the problems with these XAI techniques, which are very successful in the computer vision domain but seemingly unreliable for medical diagnostics, and also to overcome these problems by creating a novel technique which takes into account the domain-specific knowledge I will have gathered during my analysis and communication with clinicians.