Transfer Learning for Mortality Prediction in Non-Small Cell Lung Cancer with Low-Resolution Histopathology Slide Snapshots.

Summary

Transfer learning from high-resolution histopathology scans significantly improves the accuracy of mortality predictions for patients with non-small cell lung cancer when using only low-resolution slide snapshots. These findings enable the development of more effective prognostic tools in clinical settings where high-resolution imaging is unavailable.

Key result

Transfer learning from high-resolution scans achieved significantly better mortality prediction performance in non-small cell lung cancer than models without transfer learning or those using general domain images.

Abstract

High-resolution whole slide image scans of histopathology slides have been widely used in recent years for prediction in cancer. However, in some cases, clinical informatics practitioners may only have access to low-resolution snapshots of histopathology slides, not high-resolution scans. We evaluated strategies for training neural network prognostic models in non-small cell lung cancer (NSCLC) based on low-resolution snapshots, using data from the Veterans Affairs Precision Oncology Data Repository. We compared strategies without transfer learning, with transfer learning from general domain images, and with transfer learning from publicly available high-resolution histopathology scans. We found transfer learning from high-resolution scans achieved significantly better performance than other strategies. Our contribution provides a foundation for future development of prognostic models in NSCLC that incorporate data from low-resolution pathology slide snapshots alongside known clinical predictors.

Key figure