Abstract
1427
Objectives: The channelized Hoteling observer (CHO) has been the most widely used as substitute for human observers in signal-location-known tasks in nuclear medicine imaging research. However, when applied in scanning mode for clinically more realistic signal-location-unknown tasks, training complexity increases substantially. In addition, in cases where absolute rankings are important, CHO must be degraded, e.g., by the addition of an internal noise model that must be calibrated for different tasks. In this work, we report a novel and computationally efficient anthropomorphic model observer based on deep learning (DeepAMO) that models the diagnostic process of a nuclear medicine physician in a clinically realistic detection task.
Methods: The proposed observer models a stylized decision process of a reader performing a detection task using a 3D image volume. The input to the proposed model observer is a composite image consisting of 3 sets of 2D slices from orthogonal views, typical of that used to view 3D volumes in clinical practice. First, the images from each 3 orientations (coronal, sagittal and transaxial) are grouped into all possible triads (sets of 3 adjacent slices). The triads are sent to a shared segmentation network trained to segment defects in triads of slices. The output segmentation masks for these triads are summed in the slicing direction to form a 2D image giving the projection of the defect in that orientation. These 3 projections then serve as the input to a classification network. The entire network is trained using human observer rating values so that the output, when applied to an input image volume, is a rating value designed to mimic the performance of human observers. The proposed model observer was trained in two stages. First, we trained a segmentation network to segment defect masks in individual triads. In stage 2, we trained the whole model observer network starting with the trained parameters of the segmentation network from stage 1. In this work we implemented and evaluated the method in the context of Tc-99m DMSA pediatric renal SPECT. The data was generated from a family of digital phantoms that realistically models internal anatomy, body morphometry and radionuclide uptake. A previously validated simulation method was used to generate the projections. Images were reconstructed using a clinical reconstruction protocol. Human observer rating data were generated by a trained graduate student. A total of 384 of the composite images described above were used. Among those, 192 were used for training (fine-tuning) and 96 for validation; the rest was used for initial training of the human observer.
Results: The histograms of rating values generated by DeepAMO using the validation data were in good agreement with those from the human observers. The AUCs for the human observer and proposed model observer were 0.933 (CI95%= [0.995,0.881]) and 0.975 (CI95% = [1.00,0.949]), respectively. Bootstrapping and nonparametric analysis were used to compute 95% confidence intervals for the AUC values. Conclusions: Results to-date show that the proposed framework is able to mimic rating values and performance of a human observer in signal-position-unknown defect detection in a 3D image dataset.