Keywords: Liver MR Imaging, Hand-Crafted features, CNN, Classification
Hepatobiliary phase (HBP) MRI using Gd-EOB-DTPA contrast improves detection and characterization of focal hepatic lesions due to increased contrast between liver parenchyma and non-hepatocellular lesions. However, adequate hepatocellular contrast enhancement (HCE) may not be achieved. The purpose of this study was to develop a machine learning approach for classifying HBP images as having adequate or inadequate HCE for detecting focal hepatic lesions. The data comprised 1201 T1w 3D HBP images from 406 patients who underwent Gd-EOB-DTPA-enhanced liver MRI. Each image was annotated by radiologists as having adequate or inadequate HCE for detecting focal hepatic lesions. The complete analysis consisted of three parts: image preprocessing to remove bias, handcrafted feature extraction and image classification. Two types of features, intensity separation and topological structure, were captured by Gaussian mixture model and Euler characteristic curves, respectively. 826 HBP images were randomly selected for training and 375 images for validation. Classification was performed using support vector machines and handcrafted performance was compared to a convolutional neural network (CNN). Performance across random subsets of the training dataset were also compared. With complete training data, AUCs were 0.880 (95% CI, 0.84-0.924) for handcrafted features and 0.919 (95% CI, 0.886-0.951) for handcrafted and CNN-learned features. Training on 100 images yielded AUCs of 0.847 (95% CI, 0.805-0.892) for handcrafted features and 0.531 (95% CI, 0.468-0.595) for CNN-learned features. Training on 400 images yielded AUCs of 0.884 (95% CI, 0.847-0.925) for handcrafted features and 0.877 (95% CI, 0.838-0.913) for CNN-learned features. Handcrafted features outperformed the CNN when training on fewer examples. Increases in training sample size improved both handcrafted and CNN performance. The CNN was more sensitive to changes in training sample size.