Improving CT Image Segmentation Accuracy Using StyleGAN Driven Data Augmentation

TL;DR

This paper introduces a StyleGAN-based data augmentation method that enhances CT image segmentation accuracy by reducing texture discrepancies between small annotated datasets and large unannotated datasets, leading to improved deep learning performance.

Contribution

The paper proposes a novel StyleGAN-driven augmentation technique that combines style transfer with small annotated datasets to improve segmentation accuracy on large, unannotated datasets.

Findings

Significant improvement in segmentation accuracy using the proposed method

Effective elimination of texture differences between datasets

Enhanced training of U-Net with augmented data

Abstract

Medical Image Segmentation is a useful application for medical image analysis including detecting diseases and abnormalities in imaging modalities such as MRI, CT etc. Deep learning has proven to be promising for this task but usually has a low accuracy because of the lack of appropriate publicly available annotated or segmented medical datasets. In addition, the datasets that are available may have a different texture because of different dosage values or scanner properties than the images that need to be segmented. This paper presents a StyleGAN-driven approach for segmenting publicly available large medical datasets by using readily available extremely small annotated datasets in similar modalities. The approach involves augmenting the small segmented dataset and eliminating texture differences between the two datasets. The dataset is augmented by being passed through six different StyleGANs that are trained on six different style images taken from the large non-annotated dataset we want to segment. Specifically, style transfer is used to augment the training dataset. The annotations of the training dataset are hence combined with the textures of the non-annotated dataset to generate new anatomically sound images. The augmented dataset is then used to train a U-Net segmentation network which displays a significant improvement in the segmentation accuracy in segmenting the large non-annotated dataset.