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

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 $$2\times $$ 2 × 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 $$4\times $$ 4 × 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.

Deep-learning-based 3D super-resolution MRI radiomics model: superior predictive performance in preoperative T-staging of rectal cancer

“…Fan et al [39] explored the usefulness of the 2D SR neural network for resolution improvement and image-based diagnosis. Farias et al [24] found that generative adversarial network (GAN) SR increased the robustness of the most important radiomics features. These two proof-of-concept studies have revealed encouraging potential to further apply DL-SR to the practice of radiomics analysis, while validation for clinical utility by workflow is still warranted.…”

“…In RC, radiomics has attained impressive performance in different oncological scenarios, including evaluating tumor biological behaviors [20], assessing treatment response [21,22], and predicting prognosis [23]. Despite these advances, radiomics features tend to be affected by anisotropic resolution and low voxel statistics in current medical imaging [24]. To enhance the robustness and stability of radiomics models, it is considerable to tackle these limitations by applying higher-resolution images in model construction.…”

“…As for its application in MRI examinations, Masutani et al [27] proposed a novel convolutional neural network (CNN) model which optimized high-frequency spatial detail in short-axis cardiac MRI imaging; in the work named SCSRN, Niaz et al [28] reported that DL-SR algorithms improved the image quality for brain morphology diagnosis. Apart from the good stability and reliability the recovered SR images showed in the multi-spaceladder, DL-SR has attracted attention in medical imaging because the radiomics features extracted from SR images were quantitatively proven to be remarkably reproducible and robust [24]. However, none of the current radiomics researches have utilized the DL-SR technique in the discovery of a specific radiomics biomarker for clinical use.…”

Radiomics/Radiogenomics in Lung Cancer: Basic Principles and Initial Clinical Results

“…A successful ML application is primarily dependent on a solid, clinically meaningful formulation of the clinical use case. It is of paramount importance when considering different use cases to weigh the final decision not only on their clinical soundness, but instead on a realistic consideration of an adequate amount of data required for efficient training of the machine learning algorithm to produce a robust model that will be accurate and general [ 6 ]. The latter is certainly true when considering input data such as the medical images that are dynamic and change often based on information representation and lack of universally adopted standardised acquisition protocols [ 7 ].…”

Data Augmentation Based on Generative Adversarial Networks to Improve Stage Classification of Chronic Kidney Disease

“…Frid-Adar et al trained a GAN model to synthesis medical images to improve the performance of CNN in liver lesion classification [29]. Farias et al proposed that the GAN-CIRCLE framework generates images with better texture and sharper edges around the lesion [30]. Nie et al proposed a GAN with a 3D FCN structure to solve the tasks of generating CT from MRI and generating 7 Tesla (7T) MRI from 3 Tesla (3T) MRI images [31].…”