Radiomics-guided deep neural networks stratify lung adenocarcinoma prognosis from CT scans

Hwan-Ho Cho,Eunjin Kim,Ho Yun Lee,Junmo Kwon,Geewon Lee,Hyunjin Park,Jonghoon Kim

doi:10.1038/s42003-021-02814-7

Abstract

Deep learning (DL) is a breakthrough technology for medical imaging with high sample size requirements and interpretability issues. Using a pretrained DL model through a radiomics-guided approach, we propose a methodology for stratifying the prognosis of lung adenocarcinomas based on pretreatment CT. Our approach allows us to apply DL with smaller sample size requirements and enhanced interpretability. Baseline radiomics and DL models for the prognosis of lung adenocarcinomas were developed and tested using local (n = 617) cohort. The DL models were further tested in an external validation (n = 70) cohort. The local cohort was divided into training and test cohorts. A radiomics risk score (RRS) was developed using Cox-LASSO. Three pretrained DL networks derived from natural images were used to extract the DL features. The features were further guided using radiomics by retaining those DL features whose correlations with the radiomics features were high and Bonferroni-corrected p-values were low. The retained DL features were subject to a Cox-LASSO when constructing DL risk scores (DRS). The risk groups stratified by the RRS and DRS showed a significant difference in training, testing, and validation cohorts. The DL features were interpreted using existing radiomics features, and the texture features explained the DL features well.

Highlights

Deep learning (DL) is a breakthrough technology for medical imaging with high sample size requirements and interpretability issues
Three pretrained DL networks were used as feature extractors and the extracted DL features were subjected to correlation analysis with radiomics features to retain important radiomics-guided features
The DL risk scores (DRS) models were compared with radiomics risk score (RRS) and important DL features were interpreted with radiomics features

Summary

Introduction

Deep learning (DL) is a breakthrough technology for medical imaging with high sample size requirements and interpretability issues. Baseline radiomics and DL models for the prognosis of lung adenocarcinomas were developed and tested using local (n = 617) cohort. DL networks trained to recognize objects in natural images can be effectively applied to medical imaging because abstract high-level information is common between natural and medical images[42]. Another major drawback of DL is the difficulty of interpretation. The approaches highlight important regions contributing to the DL network and have a weak link to the underlying biology This issue becomes more critical in medical imaging because the decisions of the algorithms require well-rooted explanations for use in clinical practice[47]. Features derived from DL models are difficult to interpret and linking them to features that are more interpretable

Objectives

Methods

Results

Conclusion