Learning fuzzy clustering for SPECT/CT segmentation via convolutional neural networks

Junyu Chen, Ye Li, Licia P. Luna, Hyun W. Chung, Steven P. Rowe, Yong Du, Lilja B. Solnes, Eric C. Frey
Medical physics, 48(7), 3860-3877, 2021.

Summary

we proposed a set of novel FCM-based loss functions for semi-, unsupervised, and supervised SPECT/CT segmentation using deep neural networks. An advantage of the proposed loss functions is that they enable the ConvNets to consider both voxel intensity and semantic information in the image during the training stage. The proposed loss functions also retain the fundamental property of the conventional fuzzy clustering, where the fuzzy overlap between the channels of softmax outputs can be adjusted by a hyper-parameter in the loss function. Various experiments demonstrated that the model trained using a dataset of simulated images generalized well and led to fast and robust segmentation on both simulated and clinical SPECT/CT images.

Network architecture

Network architecture

Qualitative results

Membership functions

Segmentation accuracy

Registration accuracy (Dice score)

BibTex

@article{chen2021learning,
title={Learning fuzzy clustering for SPECT/CT segmentation via convolutional neural networks},
author={Chen, Junyu and Li, Ye and Luna, Licia P and Chung, Hyun W and Rowe, Steven P and Du, Yong and Solnes, Lilja B and Frey, Eric C},
journal={Medical physics},
volume={48},
number={7},
pages={3860--3877},
year={2021},
publisher={Wiley Online Library}
}