Rationale and objectives: To evaluate the ability of artificial neural networks (ANN) fed with radiomic signatures (RSs) extracted from multidetector computed tomography images in differentiating the histopathological grades of clear cell renal cell carcinomas (ccRCCs).
Materials and methods: The multidetector computed tomography images of 227 ccRCCs were retrospectively analyzed. For each ccRCC, 14 conventional image features (CIFs) were extracted manually by two radiologists, and 556 texture features (TFs) were extracted by a free software application, MaZda (version 4.6). The high-dimensional dataset of these RSs was reduced using the least absolute shrinkage and selection operator. Five minimum mean squared error models (minMSEMs) for predicting the ccRCC histopathological grades were constructed from the CIFs, the TFs of the corticomedullary phase images (CMP), and the TFs of the parenchyma phase (PP) images and their combinations, respectively abbreviated as CIF-minMSEM, CMP-minMSEM, PP-minMSEM, CIF+CMP-minMSEM, and CIF+PP-minMSEM. The RSs of each model were fed 30 times consecutively into an ANN for machine learning, and the predictive accuracy of each time ML was recorded for the statistical analysis.
Results: The five predictive models were constructed from 12, 19, and 10 features selected from the CIFs, the TFs of the CMP images, and that of PP images, respectively. On the basis of their accuracy across the whole cohort, the five models were ranked as follows: CIF+CMP-minMSEM (accuracy: 94.06% ± 1.14%), CIF + PP-minMSEM (accuracy: 93.32% ± 1.23%), CIF-minMSEM (accuracy: 92.26% ± 1.65%), CMP-minMSEM (accuracy: 91.76% ± 1.74%), and PP-minMSEM (accuracy: 90.89% ± 1.47%).
Conclusion: Machine learning based on ANN helped establish an optimal predictive model, and TFs contributed to the development of high accuracy predictive models. The CIF+CMP-minMSEM showed the greatest accuracy for differentiating low- and high-grade ccRCCs.
Keywords: Artificial neural networks; Clear cell renal cell carcinoma; Computed tomography; Machine learning; Radiomic signature; Texture feature.
Copyright © 2019 The Association of University Radiologists. Published by Elsevier Inc. All rights reserved.