Impact of GAN-based lesion-focused medical image super-resolution on the robustness of radiomic features

Erick Costa De Farias,Leonardo Rundo,Mauro Castelli,Evis Sala,Christian Di Noia,Changhee Han

doi:10.1038/s41598-021-00898-z

Abstract

Robust machine learning models based on radiomic features might allow for accurate diagnosis, prognosis, and medical decision-making. Unfortunately, the lack of standardized radiomic feature extraction has hampered their clinical use. Since the radiomic features tend to be affected by low voxel statistics in regions of interest, increasing the sample size would improve their robustness in clinical studies. Therefore, we propose a Generative Adversarial Network (GAN)-based lesion-focused framework for Computed Tomography (CT) image Super-Resolution (SR); for the lesion (i.e., cancer) patch-focused training, we incorporate Spatial Pyramid Pooling (SPP) into GAN-Constrained by the Identical, Residual, and Cycle Learning Ensemble (GAN-CIRCLE). At 2times SR, the proposed model achieved better perceptual quality with less blurring than the other considered state-of-the-art SR methods, while producing comparable results at 4times SR. We also evaluated the robustness of our model’s radiomic feature in terms of quantization on a different lung cancer CT dataset using Principal Component Analysis (PCA). Intriguingly, the most important radiomic features in our PCA-based analysis were the most robust features extracted on the GAN-super-resolved images. These achievements pave the way for the application of GAN-based image Super-Resolution techniques for studies of radiomics for robust biomarker discovery.

Highlights

Robust machine learning models based on radiomic features might allow for accurate diagnosis, prognosis, and medical decision-making
We found that the most important radiomic features in our Principal Component Analysis (PCA)-based examination were the most robust features extracted on the Generative Adversarial Network (GAN)-super-resolved images
The GAN-SR technique, as well as the Bicubic interpolation, achieved moderate to good robustness for Gray-Level Run Length Matrix (GLRLM) LongRunLowGrayLevelEmphasis and Gray-Level Dependence Matrix (GLDM) DependenceEntropy, while the features extracted from the Original images resulted in excellent robustness

Summary

Introduction

Robust machine learning models based on radiomic features might allow for accurate diagnosis, prognosis, and medical decision-making. The most important radiomic features in our PCAbased analysis were the most robust features extracted on the GAN-super-resolved images. Robust machine learning models based on large-scale radiomic features might allow for accurate diagnosis, prognosis, and medical decision-making; thoroughly considering the whole radiomic processes is essential to obtain these reliable models. Despite the potential of radiomics, high quantitative feature variability across different software implementations has hampered its clinical u se[2,3] This phenomenon derives from the lack of standardized definitions and extraction of radiomic features with validated reference values. As the paper’s authors highlighted, image features still need to be robust against differences in acquisition, reconstruction, and segmentation to ensure reproducibility For this reason, recent studies have investigated the robustness of radiomic features in several scenarios and applications using heterogeneous datasets. Along with GAN-based medical image SR, novel approaches based on progressive GANs20 and attention mechanisms[21] have been recently applied to video SR

Objectives

Methods

Results

Conclusion