Supervised and unsupervised loss functions for ConvNet image segmentation based on the classical FCM objective function.
This is a Python implementation (TensorFlow and Pytorch) of my paper:
PDF: arXiv pre-print
We present semi-, un-, and supervised loss functions based on the objective function of a Fuzzy C-means algorithm. The unsupervised loss function does not depend on the ground truth label map, enabling the unsupervised (self-supervised) training of a neural network. Combined with the proposed supervised loss, we form a semi-supervised loss function. This loss function leverages both intensity distribution and ground-truth labels, which improved our segmentation network's generalizability. Our paper showed that a ConvNet trained with purely simulation images can still yield usable segmentation for clinical images (unseen images from training dataset).
Example predictions obtained using the unsupervised RFCM loss (i.e., the network was trained using images without ground truth labels):
Note that the networks were trained using purely simulated images and tested on the "unseen" clinical patient images.
If you find this code is useful in your research, please consider to cite:
@article{https://doi.org/10.1002/mp.14903,
author = {Chen, Junyu and Li, Ye and Luna, Licia P. and Chung, Hyun Woo and Rowe, Steven P. and Du, Yong and Solnes, Lilja B. and Frey, Eric C.},
title = {Learning Fuzzy Clustering for SPECT/CT Segmentation via Convolutional Neural Networks},
journal = {Medical Physics},
volume = {n/a},
number = {n/a},
pages = {},
doi = {https://doi.org/10.1002/mp.14903},
url = {https://aapm.onlinelibrary.wiley.com/doi/abs/10.1002/mp.14903},
eprint = {https://aapm.onlinelibrary.wiley.com/doi/pdf/10.1002/mp.14903}}