Using deep learning to predict tumor mutational burden from scans of H&E-stained multicenter slides of lung squamous cell carcinoma

Purpose: A high tumor mutational burden (TMB) is a promising biomarker for identifying lung squamous cell carcinoma (SqCC) patients who are more likely to benefit from risky but potentially highly beneficial immunotherapy, but it is not available in most clinics. It has been shown that it is possible to predict TMB from standard-of-care cancer histology slides using deep learning for various cancer sites. Our goal is to build a model that can do this specifically for lung SqCC and to validate it on a held-out test set from centers on which the model was not trained.Approach: We obtained scans of diagnostic slides from 50 lung SqCC patients, with one slide per-patient, from 35 different centers. We held out 20 slides from 15 centers for testing and used the rest for training and validation, ensuring that no center was represented in more than one set. Using transfer learning, we explored several neural network architectures and training parameters to choose an optimal model.Results: Using the training and validation sets, we found the optimal model to be VGG16. The per-patient AUC for this model on the held-out test set was 0.65, with an accuracy of 65%, true positive rate of 77%, and true negative rate of 43%.Conclusions: A deep learning model can predict TMB from scans of H & E-stained slides of lung SqCC resections on an independent test set containing images only from centers on which the model was not trained. With further development and external validation, such a system can act as an alternative to traditional genetic sequencing for patients with SqCC; this will help physicians determine, with more accuracy, whether patients should be given immunotherapy. This will more effectively give access to immunotherapy drugs to those who need them and help spare others the toxicities associated with them.