CNN-based detection of pediatric lymphoma on whole body [18F]FDG-PET/MRI.

Abstract

We assessed the performance of a deep convolutional neural network (CNN) in detecting pediatric lymphoma lesions on [¹⁸F]FDG-PET/MRI. We evaluated CNN's sensitivity, specificity, percentage agreement, and processing time compared to the interpretations of a pediatric radiologist and a second-year radiology resident. In this retrospective study, a CNN was trained on annotated [¹⁸F]FDG-PET/MRI scans from 53 pediatric lymphoma patients and tested on 30 additional scans. The CNN and two human readers recorded the presence of lesions in five anatomical regions. An additional pediatric radiologist and a nuclear medicine physician determined the reference standard. The sensitivity and specificity of the CNN were compared with those of human readers using the McNemar test, and the detection time of the CNN and human readers was compared using the Wilcoxon signed-rank test. The CNN demonstrated higher sensitivity (84.6%) and specificity (93.7%) than the radiology resident (69.2%, P=0.023; 81.5%, P<0.001), but lower than the pediatric radiologist (98.7%, P<0.001; 99.5%, P<0.001). The CNN achieved 83% agreement with the reference standard (95% CI: 79%-87%), higher than the resident's 63% (95% CI: 59%-69%) but lower than the pediatric radiologist's 94% (95% CI: 92%-97%). The median values and interquartile ranges for the time taken (in minutes) were 4 (3, 5) for the CNN, 8 (7, 10) for the pediatric radiologist, and 15 (9, 20) for the radiology resident. The sensitivity, specificity, and percentage agreement of the CNN were higher than those of a radiology resident but lower than those of a pediatric radiologist. The CNN readout was significantly faster compared to both human readers.