Author: Hongyi Jiang, Fang-Fang Yin 👨🔬
Affiliation: Duke University, Medical Physics Graduate Program, Duke Kunshan University 🌍
Purpose:
Imaging moving tissues using PET-CT can be difficult. Separating signal into phases during construction reduces signal count and increases influence of noise. Algorithms that use signal from multiple phases can access more information and may be able to bypass these issues.
This study attempts to develop a deep neural network (DNN) that uses 4D sinograms as input directly and automatically produces lesion segmentations.
Methods:
In a simulation study, Xcat software generated 50 sets of 4D phantoms using randomized body shape parameters. Each 4D phantoms has 10 phases, 576 Z-layers along vertical axis, 8 different sets of lesions, and 4 different sets of shifts and rotations. Thus creating almost 10 million different 2D planes, and each plane can be used in a dataset. These phantoms were scanned using MatLab to produce 50 pairs of 4D PET-CT sinograms. Datasets from 5 phantoms were reserved for validation, and the rest were used for training.
The DNN is consisted by first a U-net, then two fully connected layers and a filtered back projection function, and finally by another U-net. In each iteration of training, a random pair of 2D PET-CT sinogram was chosen. The 2D sinogram pairs from neighboring phases and Z-layers were also inputted. Thus the input included 9 pairs of PET-CT sinograms. 2D lesion segmentation image corresponding to central sinogram pair was used as ground truth. This segmentation was produced from a 2D plane of the phantom. During validation, the trained DNN was used to process validation datasets. Results were compared with ground truths using Dice Index (DSC).
Results:
When DNN output is quantitatively compared with ground truth, the resulting DSC value was 0.9995-+0.00029.
Conclusion:
The DNN developed was able to segment lesions, including some highly unobvious ones. Quantitative comparison of this DNN with conventional convolution networks can be further investigated.