ESTRO 2025 - Abstract Book

S3239

Physics - Intra-fraction motion management and real-time adaptive radiotherapy

ESTRO 2025

For the patient data, we select difficult examples of heatmaps generated on over-saturated areas (Figure 2). Only GBP and DeepLIFT avoid areas with no information, while PatternAttribution keeps highlighting the upper half of the predicted tumor’s bounding box and LRP focuses on the center region of the bounding box, which includes overexposed areas.

Conclusion: Our findings suggest that GBP and DeepLIFT demonstrate a reliable and consistent behavior across all patients and the phantom, LRP showing good performance only in the phantom setting. We argue that GBP and DeepLIFT can be used out-of-the-box to explain deep learning-based tracking models for SBRT using VMAT. Further investigation is needed to develop a robust measure of model’s reliability, based on generated heatmaps, during treatment delivery.

Keywords: lung tumor tracking, deep learning, explainable AI

References: 1) C. Hazelaar et al., Markerless positional verification using template matching and triangulation of kV images acquired during irradiation for lung tumors treated in breath-hold 2) J. T. Springenberg et al., Striving for simplicity: The all convolutional net 3) S. Bach et al., On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation 4) A. Shrikumaret al., Learning important features through propagating activation differences 5) P.-J. Kindermans et al., Learning how to explain neural networks: Patternnet and patternattribution