Torwards Trustworthy Machine Learning based systems: Evaluating breast cancer predictions interpretability using Human Centered Machine Learning and UX Techniques

Abstract

Although the use of Machine Learning techniques has been widely used in the literature in order to predict breast cancer. The focus of these works has been to improve the performance of classification algorithms for greater diagnostic accuracy. However, for these classification models to be used in a real environment, such as a cancer diagnosis assistance system in an oncology institution, in addition to high performance, models must also offer predictions that are easy to understand by radiologists who make the final diagnosis. In this work, we evaluate the level of trust from users in an AI-based system for breast cancer identification. This system uses computer vision and Deep Learning (DL) techniques to classify breast mammography and identify abnormalities associated with lumps or cancer tumors. The evaluation performed in this work focuses on the interpretability of the system and the explanations that are shown to users. To evaluate the interpretability of the model’s predictions, AI-based systems evaluation techniques from the Human-Centered Machine Learning (HCML) field were used, as well as classic usability and user experience (UX) techniques.

The results obtained show that users’ trust is related to the presentation of the explanations, that is, to how the system UI displays the predictions and shows the zones of the images used to calculate the predictions. In this sense, it was also possible to observe that the classic techniques of usability and UX have a relationship with the level of trust perceived by the users, which was measured with HCML evaluation techniques.