Unsupervised non-small cell lung cancer tumor segmentation using cycled generative adversarial network with similarity-based discriminator

Abstract

Background

Tumor segmentation is crucial for lung disease diagnosis and treatment. Most existing deep learning-based automatic segmentation methods rely on manually annotated data for network training.

Purpose

This study aims to develop an unsupervised tumor segmentation network smic-GAN by using a similarity-driven generative adversarial network trained with cycle strategy. The proposed method does not rely on any manual annotations and thus reduce the training data preparation workload.

Methods

A total of 609 CT scans of lung cancer patients are collected, of which 504 are used for training, 35 for validation, and 70 for testing. Smic-GAN is developed and trained to transform lung CT slices with tumors into synthetic images without tumors. Residual images are obtained by subtracting synthetic images from original CT slices. Thresholding, 3D median filtering, morphological erosion, and dilation operations are implemented to generate binary tumor masks from the residual images. Dice similarity, positive predictive value (PPV), sensitivity (SEN), 95% Hausdorff distance (HD95) and average surface distance (ASD) are used to evaluate the accuracy of tumor contouring.

Results

The smic-GAN method achieved a performance comparable to two supervised methods UNet and Incre-MRRN, and outperformed unsupervised cycle-GAN. The Dice value for smic-GAN is significantly better than cycle-GAN (74.5% $\pm$ 11.2% vs. 69.1% $\pm$ 16.0%, p < 0.05). The PPV for smic-GAN, UNet, and Incre-MRRN are 83.8% $\pm$ 21.5%,75.1% $\pm$ 19.7%, and 78.2% $\pm$ 16.6% respectively. The HD95 are 10.3 $\pm$ 7.7, 14.5 $\pm$ 14.6 and 6.2 $\pm$ 4.0 mm, respectively. The ASD are 3.7 $\pm$ 2.7, 4.8 $\pm$ 3.8, and 2.4 $\pm$ 1.8 mm, respectively.

Conclusion

The proposed smic-GAN performs comparably to the existing supervised methods UNet and Incre-MRRN. It does not rely on any manual annotations and can reduce the workload of training data preparation. It can also provide a good start for manual annotation in the training of supervised networks.