Variability in reference levels for Deauville classifications applied to lymphoma patients examined with 18F-FDG-PET/CT

Journal article, 2017

Aim: To develop and validate a convolutional neural network (CNN) based method for automated quantification of reference levels in liver and mediastinum (blood pool) for the Deauville therapy response classification using FDG-PET/CT in lymphoma patients.
Methods: CNNs were trained to segment the liver and the mediastinum, defined as the thoracic part of the aorta, in CT images from 81 consecutive lymphoma patients, who had undergone FDG-PET/CT examinations. Trained image readers segmented the liver and aorta manually in each of the CT images and these segmentations together with the CT images were used to train the CNN. After the training process, the CNN method was applied to a separate validation group consisting of six consecutive lymphoma patients (17-82 years, 3 female). First, the liver and mediastinum were automatically segmented in the CT images. Second, voxels in the corresponding FDG-PET images, which were localized in the liver and mediastinum, were selected and the median standard uptake value (SUV) was calculated. The CNN based analysis was compared to corresponding manual segmentations by two experienced radiologists. The Dice index was used to analyse the overlap between the segmentations by the CNN and the two radiologists. A Dice index of 1.00 indicates perfect matching.
Results: The mean Dice indices for the comparison between CNN based liver segmentations and those of the two radiologists in the validation group were 0.95 and 0.95. A corresponding comparison between the two radiologists also resulted in a Dice index of 0.95. The mean liver volumes were 1,752ml, 1,757ml and 1,768ml for the CNN and two radiologists, respectively. The median SUV for the liver was on average 1.8 and the differences between median SUV based on CNN and manual segmentations were less or equal to 0.1. The mean Dice indices for the mediastinum were 0.80, 0.83 (CNN vs radiologists) and 0.86 (comparing the two radiologists). The mean mediastinum (aorta) volumes were 147ml, 140ml and 125ml for the CNN and
two radiologists, respectively. The median SUV for the mediastinum was on average 1.4 and the differences between median SUV based on CNN and manual segmentations were less or equal to 0.14.
Conclusion: A CNN based method for automated quantification of reference levels in liver and mediastinum show good agreement with results obtained by experienced radiologists, who manually segmented the CT images. This is a first and promising step towards a completely objective treatment response evaluation in patients with lymphoma based on FDG-PET/CT.