An Explainable AI Framework to Support Art Therapy.

Art therapy uses the analysis of drawings for therapeutic purposes. Often, drawings can be a more effective means of communication, especially when children are involved. However, the most important first step for parents is to be aware of potential alarm signals in their children’s paintings and to seek dialogue with the child or the assistance of an art therapist. A technological solution to facilitate this first step is currently missing. In this thesis, we therefore propose an explainable, AI-based framework to support art therapy. The proposed system analyses children’s drawings automatically and uses a YOLO-based (“You Only Look Once”) pipeline to localize and classify objects in these drawings. Our system reports its findings not in the traditional way of a single bounding box and percentages but uses plain English to reach laypersons as well. Taking a capability-aware approach, our system also avoids reporting numbers suggesting high confidence in the precision of the black box prediction process. Instead, the results are presented in a verifiable manner to parents that allow for an informed decision when considering whether to contact an art therapist. In this thesis, we not only review related work and present our framework, but we also discuss ethical considerations arising from using black box inference in a medical context.