CNN as model observer in a liver lesion detection task for x-ray computed tomography: A phantom study.

Abstract

The purpose of this study was the evaluation of anthropomorphic model observers trained with neural networks for the prediction of a human observer's performance. To simulate liver lesions, a phantom with contrast targets (acrylic spheres, varying diameters, +30HU) was repeatedly scanned on a computed tomography scanner. Image data labeled with confidence ratings assessed in a reader study for a detection task of liver lesions were used to build several anthropomorphic model observers. Models were trained with images reconstructed with iterative reconstruction and evaluated with images reconstructed with filtered backprojection. A neural network, based on softmax regression (SR-MO), and convolutional neural networks (CNN-MO) were used to predict the performance of a human observer and compared to a channelized Hotelling observer [with Gabor channels and internal channel noise (CHOi)]. Model observers were evaluated by a receiver operating characteristic curve analysis and compared to the results in the reader study. Two strategies were used to train the SR-MO and CNN-MO: A) building a separate model for each lesion size; B) building one model that was applied to lesions of all sizes. All tested model observers and the human observer were highly correlated at each lesion size and dose level. With strategy A, Pearson's product-moment correlation coefficients r were 0.926 (95% confidence interval (CI): 0.679-0.985) for SR-MO and 0.979 (95% CI: 0.902-0.996) for CNN-MO. With strategy B, r was 0.860 (95% CI: 0.454-0.970) for SR-MO and 0.918 (95% CI: 0.651-0.983) for CNN-MO. For CHOi, r was 0.945 (95% CI: 0.755-0.989). With strategy A, mean absolute percentage differences (MAPD) between the model observers and the human observer were 3.7% for SR-MO and 1.2% for CNN-MO. With strategy B, MAPD were 3.7% for SR-MO and 3.0% for CNN-MO. For the CHOi the MAPD was 2.2%. Convolutional neural network model observers can accurately predict the performance of a human observer for all lesion sizes and dose levels in the evaluated signal detection task.