ExplainabilityTransparencyTrustworthinesscs.CV
Human-centered XAI for Burn Depth Characterization
Maxwell J. Jacobson, Daniela Chanci Arrubla, Maria Romeo Tricas, Gayle Gordillo, Yexiang Xue, Chandan Sen, Juan Wachs
October 24, 2022
2 citations
Key Takeaways
-
Expert knowledge improves AI reliability: The framework improved burn depth classification accuracy from 88% to 94% by using LIME explanations to integrate burn specialists' expertise into the model, demonstrating that human-in-the-loop approaches can catch what algorithms miss on their own.
-
Explainability revealed actionable insights: The explanation system identified specific texture features (Gray Level Co-occurrence Matrix features) that the initial model overlooked, showing that making AI decisions transparent can lead to concrete technical improvements rather than just transparency for its own sake.
-
Real-world medical impact: With 1.25 million Americans treated annually for burn injuries, improving classification accuracy by 6 percentage points could meaningfully affect treatment decisions and patient outcomes in a high-stakes clinical domain.
-
Validation matters for medical AI: Rather than accepting the model's initial performance, the researchers validated their findings against expert judgment and real data, illustrating why responsible AI development in healthcare requires rigorous verification beyond standard accuracy metrics.
Abstract
Approximately 1.25 million people in the United States are treated each year for burn injuries. Precise burn injury classification is an important aspect of the medical AI field. In this work, we propose an explainable human-in-the-loop framework for improving burn ultrasound classification models. Our framework leverages an explanation system based on the LIME classification explainer to corroborate and integrate a burn expert's knowledge -- suggesting new features and ensuring the validity of the model. Using this framework, we discover that B-mode ultrasound classifiers can be enhanced by supplying textural features. More specifically, we confirm that texture features based on the Gray Level Co-occurance Matrix (GLCM) of ultrasound frames can increase the accuracy of transfer learned burn depth classifiers. We test our hypothesis on real data from porcine subjects. We show improvements in the accuracy of burn depth classification -- from ~88% to ~94% -- once modified according to our framework.