Abstract
1409
Objectives: To evaluate the repeatability of using a convolutional neural network to perform classification and location prediction of PET foci by using as inputs to the network CT volumes with different reconstruction methods.
Methods: Nineteen patients underwent 18F-FDG PET/CT scans on both Biograph mCT and Biograph Vision (Siemens Healthineers) systems. For each scan, the CT volume was reconstructed with a method A (convolution kernel B19f-mCT, B30f-Biograph Vision; pixel spacing 1.52 mm; slice thickness 2 mm) and a method B (convolution kernel B31f-mCT, l31f\3-Vision; pixel spacing 0.98 mm; slice thickness 3 mm), with PET reconstruction on Biograph Vision (PSF + TOF 4i5s, voxel size: 1.65 x 1.65 x 2 or 3.3 x 3.3 x 2 mm, uptake time:~106 min) and mCT (PSF + TOF 3i21s, voxel size: 2.03 x 2.03 x 2 mm, uptake time: ~ 65 min). PET foci for each scan were automatically identified using SUVpeak with the PERCIST liver threshold and segmented using a local maximum threshold of 42%; the reference region for the liver was determined using the CT volume reconstructed with method B. A previously published convolutional neural network [1] was used to classify and localize the PET foci. There were two network designs; the first takes as an input both PET/CT volumes to classify the foci as non-suspicious or suspicious for cancer and predict location, while the second network design takes the CT volume only to predict the location of the foci. The 137 possible locations of foci follow the established ICD-10 naming convention, specifying the body part, region, sub-region. Laterality was also taken into consideration. For each scan, the CT volumes reconstructed using methods A and B were separately input to the network with and without the PET volume to compare the outputs of the network.
Results: A total of 1,819 foci were segmented from the 38 scans. The classification prediction output was consistent between the two CT reconstructions in 98.7% of the foci. The overall location : [body part, region, sub-region, laterality] prediction was consistent in 94.6% : [99.2, 96.7, 96.3, 98.1]% and 92.6% : [98.4, 95.1, 94.6, 97.4]% of the foci when using both PET/CT and CT only as input, respectively. Per patient-scan, the classification prediction consistency between reconstructions on both scanners [mCT, Vision] was 98.4% [98.2%, 98.6%], and the location prediction consistencies were 94.0% [92.3%, 95.7%] when using PET and CT inputs and 92.0% [91.3%, 92.8%] for CT only. For both cases, using the network with the CT volume only yielded less consistent results for location prediction.
Conclusions: This study provides evidence that the designed convolutional neural network is robust in predicting foci characterization irrespective of the CT reconstruction method. For location prediction, a good agreement is observed at body part level; however, a significant difference is observed in predicting the complete ICD-10 location of the foci. The PET/CT network has less variability compared to network with only CT as an input. On the scanner with higher PET sensitivity [Biograph Vision], the differences due to varying CT reconstruction methods for location prediction are reduced, but patients were scanned at a later time; further study is required that does not have potential uptake time bias. Thus, such parameters need to be taken into consideration during design and training of neural networks for such tasks. Inclusion of CT images with different reconstruction kernels during training of the network may help to improve repeatability and performance of networks designed for this task. References: [1] Sibille, L ., Seifert R., Avramovic N., et.al., “18F-FDG PET/CT Uptake Classification in Lymphoma and Lung Cancer by Using Deep Convolutional Neural Networks”, Radiology - https://pubs.rsna.org/doi/abs/10.1148/radiol.2019191114