Prediction of lesion shrinkage using CT imaging with radiomic and deep learning approaches.

Abstract

e14592 Background: Radiomic features derived from CT scans have shown promise in predicting treatment response (Sun et al 2018, and others). We carried out a proof-of-concept study to investigate the use of CT images to predict lesion-level response. Methods: CT images from Merck studies KEYNOTE-010 (NCT01905657) and KEYNOTE-024 (NCT02142738), were used. Data from each study were evaluated separately and split for training (80%) and validation (20%) in each study. A lesion was classified as “shrinking” if ≥30% size reduction from baseline was seen on any future scan. There were 2004 (613 shrinking vs. 1391 non-shrinking) and 588 (311 vs. 277) lesions in KN10 and KN24, respectively. 130 radiomic features were extracted, followed by random forest to predict lesion response. In addition, end-to-end deep learning was used, which predicts the response directly from ROIs of CT images. Models were trained in two ways: (1) using pre-treatment baseline (BL) only or (2) using both BL and the first post-treatment image (V1) as predictors. Finally, to evaluate the predictive power without relying on initial lesion size, size information was omitted from CT images. Results: Results from the KN10 and KN24 are summarized in Table. Conclusions: The results suggest that the BL CT images alone have little power to predict lesion response, while BL and the first post-baseline image exhibit high predictive power. Although a substantial part of the predictive power can be attributed to change in ROI size, the predictive power does exist in other aspects of CT images. Overall, the radiomic signature followed by random forest produced predictions similar to, if not better than, the deep learning approach. [Table: see text]