Author: Ashok Bhandari, Kurtis Johnson, Yu Lei, Tianzhe Li, Kyuhak Oh, Shuo Wang, Chi Zhang, Su-Min Zhou 👨🔬
Affiliation: University of Nebraska Medical Center 🌍
Purpose: The purpose of this study was to leverage neural network visualization tools to detect high-attention areas indicative of treatment failure in lung SBRT by examining correlation matrices between CT and Biological Effective Dose (BED).
Methods: We retrospectively studied 179 lung cancer patients treated with SBRT between 2007 and 2022, among which 172 patients had quality 4D planning CT. Each clinical plan was recalculated on Free-Breathing (FB) CT or Average Intensity Projection (AIP) CT, either one that was not used in the original plan calculation. We created the voxelated BED dose matrix for each plan based on their prescription and calculated four correlation matrices between CT and voxelated BED: Entropy, Spearman Rank, Jensen–Shannon Divergence (JSD) and Wasserstein Distance (WD). For each correlation matrix, we developed a 17-layer CNN model to predict treatment failure and generated a high-attention region by applying a threshold at the 95th percentile of the values within the 3D attention maps generated by Grad-CAM.
Results: With a median follow-up time of 34.4 months, the FB/AIP dataset included 125/120 non-failure cases and 54/52 failure cases. All CNN models using correlation matrices achieved 100% accuracy in predicting treatment failure on both datasets, with high-attention areas overlapping significantly more with a peritumoral region (2cm outside PTV) than PTV. The high-attention areas identified by models using WD, Spearman, JSD, and Entropy matrices overlapped with the peritumoral region at percentages of 42.3%, 25%, 25.5%, and 33.2% for the FB dataset, and 45%, 24.8%, 28.8%, and 30.8% for the AIP dataset. Overlap with the peritumoral region was significantly higher than with PTV, which had overlap percentages of 12.6%, 7.4%, 5.7%, and 9.2% for the FB dataset and 9%, 8%, 8.5%, and 15.2% for the AIP dataset
Conclusion: CNN visualization revealed high-attention areas associated with treatment failure predominantly in the peritumoral region.