ESTRO 2024 - Abstract Book

S5117

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2024

Class imbalance during training was offset by a balanced class sampling approach using TorchIO [2]. This ensured that 50% of the patches within the batch contained central voxels with tumors. Cross-entropy loss was used for training the model and the performance metric was optimized for pseudo-receiver operating characteristic-area under the curve (ROC-AUC) evaluated on class-balanced patches sampled from the validation set. During inference, the predictions were made on the entire input images based on a sliding window approach to cover the original image grid. The window size matched the patch size utilized during training while the stride was adjusted such that the adjacent predictions overlap by 50%. A stratified 5-fold cross-validation was utilized to gauge a robust estimate of the predictive performance of the model.

Results:

The model demonstrated a mean cross-validation ROC-AUC of 0.9185 (std=0.0446) with a Brier score of 0.0275 (std=0.0076) in predicting PD-L1 expression. The mean balanced accuracy was also found to be 0.8730 (std = 0.0525). Figure 2 illustrates the probability maps generated by the model for both PD-L1 negative and positive cases.

Conclusion:

Although the study necessitates rigorous validation, high-b value DWI MR could potentially be used as a noninvasive technique to predict PD-L1 expression in patients with NSCLC, a crucial factor in determining eligibility for immunotherapy.

Keywords: deep learning, prediction, pdl1

References:

[1] Paszke A, Gross S, Massa F, Lerer A, Bradbury J, Chanan G, Killeen T, Lin Z, Gimelshein N, Antiga L, Desmaison A. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems. 2019;32. [2] Pérez-García F, Sparks R, Ourselin S. TorchIO: a Python library for efficient loading, preprocessing, augmentation and patch-based sampling of medical images in deep learning. Computer Methods and Programs in Biomedicine. 2021 Sep 1;208:106236.