Deep Learning for Predicting Effect of Neoadjuvant Therapies in Non–Small Cell Lung Carcinomas With Histologic Images

Abstract

Neoadjuvant therapies are used for locally advanced non–small cell lung carcinomas, whereby pathologists histologically evaluate the effect using resected specimens. Major pathological response (MPR) has recently been used for treatment evaluation and as an economical survival surrogate; however, interobserver variability and poor reproducibility are often noted. The aim of this study was to develop a deep learning (DL) model to predict MPR from hematoxylin and eosin–stained tissue images and to validate its utility for clinical use. We collected data on 125 primary non–small cell lung carcinoma cases that were resected after neoadjuvant therapy. The cases were randomly divided into 55 for training/validation and 70 for testing. A total of 261 hematoxylin and eosin–stained slides were obtained from the maximum tumor beds, and whole slide images were prepared. We used a multiscale patch model that can adaptively weight multiple convolutional neural networks trained with different field-of-view images. We performed 3-fold cross-validation to evaluate the model. During testing, we compared the percentages of viable tumor evaluated by annotator pathologists (reviewed data), those evaluated by nonannotator pathologists (primary data), and those predicted by the DL-based model using 2-class confusion matrices and receiver operating characteristic curves and performed a survival analysis between MPR-achieved and non-MPR cases. In cross-validation, accuracy and mean F1 score were 0.859 and 0.805, respectively. During testing, accuracy and mean F1 score with reviewed data and those with primary data were 0.986, 0.985, 0.943, and 0.943, respectively. The areas under the receiver operating characteristic curve with reviewed and primary data were 0.999 and 0.978, respectively. The disease-free survival of MPR-achieved cases with reviewed and primary data was significantly better than that of the non-MPR cases (P<.001 and P=.001), and that predicted by the DL-based model was almost identical (P=.005). The DL model may support pathologist evaluations and can offer accurate determinations of MPR in patients.