Improving Feature Extraction from Histopathological Images Through A Fine-tuning ImageNet Model

Abstract

Due to lack of annotated pathological images, transfer learning has been the predominant approach in the field of digital pathology.Pre-trained neural networks based on ImageNet database are often used to extract "off the shelf" features, achieving great success in predicting tissue types, molecular features, and clinical outcomes, etc. We hypothesize that fine-tuning the pre-trained models using histopathological images could further improve feature extraction, and downstream prediction performance.We used 100,000 annotated HE image patches for colorectal cancer (CRC) to finetune a pretrained Xception model via a twostep approach.The features extracted from finetuned Xception (FTX2048) model and Imagepretrained (IMGNET2048) model were compared through: (1) tissue classification for HE images from CRC, same image type that was used for finetuning; (2) prediction of immunerelated gene expression and (3) gene mutations for lung adenocarcinoma (LUAD).Fivefold cross validation was used for model performance evaluation. The extracted features from the finetuned FTX2048 exhibited significantly higher accuracy for predicting tisue types of CRC compared to the off the shelf feature directly from Xception based on ImageNet database. Particularly, FTX2048 markedly improved the accuracy for stroma from 87% to 94%. Similarly, features from FTX2048 boosted the prediction of transcriptomic expression of immunerelated genesin LUAD. For the genes that had signigicant relationships with image fetures, the features fgrom the finetuned model imprroved the prediction for the majority of the genes. Inaddition, fetures from FTX2048 improved prediction of mutation for 5 out of 9 most frequently mutated genes in LUAD.