PCEvE: Part Contribution Evaluation Based Model Explanation for Human Figure Drawing Assessment and Beyond

Part-based explanations are more practical than pixel-level ones: The paper demonstrates that explaining AI decisions through identifiable parts of drawings (e.g., head, arms, legs) is more interpretable than traditional methods that highlight individual pixels, making it easier for clinicians and non-technical users to understand why a model made a specific diagnosis.

Shapley Values quantify which drawing features matter most: By using Shapley Values to measure each part's contribution to the model's decision, PCEvE provides a clear ranking of which elements of a human figure drawing were most influential—for example, whether the presence or quality of facial features was more important than body proportions in an autism assessment.

Explanations go beyond individual cases: Unlike existing explainability methods that only explain single predictions, PCEvE offers insights at three levels: individual sample explanations, class-level patterns (what features typically distinguish one diagnosis from another), and task-level understanding (how the model generally approaches the assessment task), providing a more comprehensive view of model behavior.

The method was rigorously tested and validated: The researchers validated PCEvE across multiple human figure drawing datasets and included controlled experiments to verify the method works as intended, plus demonstrated its applicability to other domains like car image classification, suggesting the approach has broader relevance beyond medical diagnosis.

Transparency is critical for clinical AI applications: The emphasis on explainability reflects an important principle for responsible AI in healthcare—when AI systems help diagnose conditions like autism, clinicians need to understand the reasoning to trust and appropriately use these tools in practice.