Structural tensor and frequency guided semi-supervised segmentation for medical images.

BACKGROUND

The method of semi-supervised semantic segmentation entails training with a limited number of labeled samples alongside many unlabeled samples, aiming to reduce dependence on pixel-level annotations. Most semi-supervised semantic segmentation methods primarily focus on sample augmentation in spatial dimensions to reduce the shortage of labeled samples. These methods tend to ignore the structural information of objects. In addition, frequency-domain information also supplies another perspective to evaluate information from images, which includes different properties compared to the spatial domain.

PURPOSE

In this study, we attempt to answer these two questions: (1) is it helpful to provide structural information of objects in semi-supervised semantic segmentation tasks for medical images? (2) is it more effective to evaluate the segmentation performance in the frequency domain compared to the spatial domain for semi-supervised medical image segmentation? Therefore, we seek to introduce structural and frequency information to improve the performance of semi-supervised semantic segmentation for medical images.

METHODS

We present a novel structural tensor loss (STL) to guide feature learning on the spatial domain for semi-supervised semantic segmentation. Specifically, STL utilizes the structural information encoded in the tensors to enforce the consistency of objects across spatial regions, thereby promoting more robust and accurate feature extraction. Additionally, we proposed a frequency-domain alignment loss (FAL) to enable the neural networks to learn frequency-domain information across different augmented samples. It leverages the inherent patterns present in frequency-domain representations to guide the network in capturing and aligning features across diverse augmentation variations, thereby enhancing the model's robustness for the inputting variations.

RESULTS

We conduct our experiments on three benchmark datasets, which include MRI (ACDC) for cardiac, CT (Synapse) for abdomen organs, and ultrasound image (BUSI) for breast lesion segmentation. The experimental results demonstrate that our method outperforms state-of-the-art semi-supervised approaches regarding the Dice similarity coefficient.

CONCLUSIONS

We find the proposed approach could improve the final performance of the semi-supervised medical image segmentation task. It will help reduce the need for medical image labels. Our code will are available at https://github.com/apple1986/STLFAL.