1155 - Illuminating the Black-Box: An Explainable AI-Driven Solution for Lung Radiotherapy Plan Evaluation with Initial Multi-Physician and Multi-Institution Assessment

Purpose/Objective(s): Evaluating radiotherapy (RT) plans is time-consuming and potentially subjective, requiring rounds of physician reviews for clinical acceptance. Traditional DVH-based methods fail to capture spatial dose-structure relationships, while deep-learning models lack explainability due to their "black-box" nature, raising safety concerns in clinical decision-making. An objective, interpretable, and automated evaluation solution is highly desirable. This study purposes a novel attention-gated eXplainable Artificial Intelligence (XAI) that mimics human physicians, providing clear, interpretable evaluation on treatment plans to enhance workflow efficiency and reduce plan quality variation in lung RT planning.

Materials/Methods: We retrospectively collected 80 locally advanced lung cancer patients treated by a single physician (60, 66, or 70 Gy, 2 Gy/fraction), splitting them into 56 for training, 14 for validation, and 10 for testing. Each case includes a planning CT, organ masks, both approved and unapproved plan doses if available. All data were resampled to 2.5×2.5×2.5 mm³ resolution. XAI was trained on patients of one physician to predict plan approval probability and generate dose maps suggesting improvements. Unlike traditional "black-box" models, XAI improves interpretability with attention gates - channel attention distinguishes anatomical features, while spatial attention identifies spatial dose regions affecting decisions. Clinical explainability was achieved through attention maps and validated against human observations. t-SNE projected complex features of approved or rejected plan doses into a lower-dimensional space to assess classification validity. To assess inter-physician and inter-institutional variations, we evaluated 10 additional patients from another physician in our clinic and 4 from a different institution, all with the same staging and prescription doses as the training data.

Results: Based on cross-validation, XAI achieved classification metrics of 0.86±0.05 accuracy, 0.93±0.04 sensitivity, and 0.79±0.04 specificity. Attention maps revealed “where” and “what” focused on plan rejection, aligning well with human observations. t-SNE clustering separated approved from rejected plans. For inter-physician review, XAI achieved 0.83 accuracy of plan approval, consistent with expectations seen in chart rounds and standardized scorecards that reduce quality variation in our clinic. In the inter-institutional study, while XAI classified external plans as approved, attention maps emphasized the heart, indicating differences in heart dose evaluation in lung RT across institutions.

Conclusion: Our XAI enables interpretable AI-driven plan evaluation, providing objective reasoning for approval or rejection and suggesting improvements. Preliminary results show promising performance across physicians and institutions, indicating the potential to streamline planning workflow and enhance plan quality consistency.