Bilateral Supervision Network for Semi-Supervised Medical Image Segmentation

Research Background and Motivation

Medical image segmentation is of great significance in the image analysis of anatomical structures and lesion areas, as well as in clinical diagnosis. However, existing fully supervised learning methods rely on a large amount of annotated data, and obtaining pixel-level annotated data for medical images is costly and time-consuming. To alleviate the dependency on annotated data, semi-supervised learning (SSL) methods have gradually emerged. Despite the promising results of existing SSL methods such as the mean teacher (MT) framework, there are still many limitations. Therefore, this study proposes a Bilateral Supervision Network (BSNet) to better utilize unlabeled samples and improve the performance of semi-supervised medical image segmentation.

Source

This paper was written by Along He, Tao Li, Juncheng Yan, Kai Wang, and Huazhu Fu. The authors are from Tianjin University Key Laboratory of Network and Data Security Technology, Nankai University School of Computer Science, Haihe Laboratory of iTAI, Singapore Institute of High Performance Computing (IHPC), and the Agency for Science, Technology and Research (A*STAR). The paper was published in the May 2024 issue of IEEE Transactions on Medical Imaging, Volume 43, Issue 5.

Research Process

The proposed Bilateral Supervision Network (BSNet) in this paper includes the following key steps:

a) Detailed Description of Research Process

1. Network Architecture Design:

   • BSNet consists of two segmentation networks fa and fb with the same structure, which can be trained on annotated data to calibrate weight deviations. Bilateral Exponential Moving Average (EMA) is used to update weights, allowing the two networks to learn from each other.
   • An additional lightweight discriminator (fd) is designed for adversarial learning to improve the reliability of pseudo-labels.

2. Bilateral Exponential Moving Average (Bilateral-EMA):

   • In the original MT framework, only EMA updates from the student model to the teacher model are considered, causing a training bottleneck. BSNet addresses this issue through Bilateral-EMA between fa and fb. The weights of fa and fb can be calibrated on annotated data to enhance the model’s learning ability.

3. Bilateral Supervision:

   • Pseudo-labels are used for bilateral supervision of the two segmentation networks. Pseudo-labels are generated mutually by the two networks and supervised each other to better utilize the class information in unlabeled data and improve the reliability of segmentation results.

4. Adversarial Learning:

   • Using the GAN mechanism, fd acts as a discriminator to differentiate between pseudo-labels and real labels. The goal is to make pseudo-labels closer to the distribution of real labels, thereby improving the supervision quality of unlabeled data.

b) Major Research Results

BSNet demonstrated excellent performance in semi-supervised segmentation tasks after experiments on multiple medical image segmentation datasets. Specific experimental results include:

1. Results on Skin Lesion Dataset:

   • When using ¹⁄₉ annotated data (100 annotated images and 800 unlabeled images), BSNet showed significant improvements in Dice and IoU scores compared to Unet and other state-of-the-art (SOTA) methods, approaching the performance of fully supervised Unet.

2. Results on Gastrointestinal Polyp Segmentation Dataset:

   • In the complex background polyp segmentation task, BSNet outperformed other SOTA methods through bilateral EMA and adversarial learning strategies.

3. Results on Retinal Vessel Segmentation Dataset:

   • In the task of fine vessel segmentation, BSNet performed excellently in IoU and Dice scores, effectively utilizing the low number of training samples.

4. Results on Prostate MRI Dataset:

   • On the Promise12 dataset, BSNet showed significant improvements in Dice and ASSD metrics compared to existing SOTA methods.

5. Results on Left Atrium Dataset:

   • On the Left Atrium MRI dataset, BSNet achieved the best performance across all evaluation metrics, verifying its effectiveness.

c) Research Conclusions and Value

BSNet successfully addresses the shortcomings of the MT method by significantly improving the performance of semi-supervised medical image segmentation through bilateral supervision and adversarial learning strategies. These results indicate that BSNet not only performs excellently on 2D medical images but also shows outstanding performance in 3D medical image segmentation. The method has important application value, especially under conditions where annotated data is scarce, as it significantly reduces dependency on a large amount of annotated data.

d) Research Highlights

• Innovations: The proposed bilateral supervision strategy and adversarial learning mechanism effectively resolve the performance bottleneck brought by MT.
• Experimental Evidence: Comprehensive experimental results on different datasets demonstrate the superior performance of BSNet.
• Applicability: The method is suitable for both large-scale lesion segmentation tasks and fine vessel segmentation tasks, showing strong generalizability.

e) Other Valuable Information

Experiments on different segmentation networks (such as Unet++ and Attention Unet) and different Backbones (such as ResNet50, ResNet101, and Transformer-based Architecture) revealed the good generalization performance of BSNet. Additionally, sensitivity analysis on different loss coefficient initialization values verified the robustness and robustness of BSNet.

Future Work

Future research can design more effective semi-supervised learning algorithms to further utilize cross-task unlabeled data to enhance SSL performance. Additionally, considering that CNNs lack long-range pixel dependency, integrating the global features of the Transformer model could better address lesion segmentation tasks in complex scenarios.