TLDR: U-Mamba2-SSL is a novel semi-supervised learning framework designed for accurate tooth and pulp segmentation in 3D CBCT scans. It employs a three-stage training strategy: self-supervised pre-training using a disruptive autoencoder, consistency regularization with input and feature perturbations on unlabeled data, and a final pseudo-labeling stage. This approach effectively leverages large amounts of unlabeled data to significantly improve segmentation performance, achieving top results in the STSR 2025 Task 1 Challenge validation set.
Accurate segmentation of teeth and pulp in Cone-Beam Computed Tomography (CBCT) scans is a critical task in modern dentistry. It plays a vital role in various clinical applications, including precise diagnosis, orthodontic procedures, and meticulous treatment and surgery planning. However, this process has traditionally been a labor-intensive endeavor, demanding extensive expertise and a significant amount of time from dental professionals. The high resolution of CBCT scans, containing a massive number of voxels, and the considerable variability across different scans make manual segmentation impractical to scale up, highlighting a pressing need for automated solutions.
This challenge has led to the development of advanced algorithms, particularly those that can effectively utilize both labeled and a large amount of unlabeled data. Semi-supervised learning (SSL) emerges as a powerful paradigm here, combining elements from both supervised and unsupervised learning to improve performance on supervised tasks by extracting latent knowledge from unlabeled data, thereby reducing the reliance on extensive, costly labels.
Introducing U-Mamba2-SSL: A Novel Framework
A new research paper introduces U-Mamba2-SSL, a novel semi-supervised learning framework designed specifically for tooth and pulp segmentation in 3D CBCT images. This innovative framework builds upon the U-Mamba2 model and employs a sophisticated multi-stage training strategy to achieve superior performance. The U-Mamba2 model itself integrates Mamba2 state space models into the U-Net architecture, enhancing its ability to capture long-range dependencies efficiently, a crucial aspect for complex 3D medical images.
The U-Mamba2-SSL framework operates through three distinct training stages:
The first stage involves **pre-training** U-Mamba2 in a self-supervised manner using a disruptive autoencoder. This autoencoder is trained to reconstruct corrupted inputs by performing three low-level reconstruction tasks: denoising (restoring original input from noisy versions), super-resolution (increasing resolution from downsampled images), and recovering masked information (filling in randomly masked regions). This initial stage allows the model to learn meaningful feature representations from all available data, both labeled and unlabeled, guiding its weights towards better starting points.
The second stage focuses on **consistency regularization training**. Here, the model is trained using a combination of supervised loss for the limited labeled data and unsupervised loss for the abundant unlabeled data. The core idea is to enforce the model to produce consistent outputs even when the input or its internal features are slightly perturbed. This involves applying strong data augmentations to the input (like median filter, Gaussian blur, noise, brightness/contrast changes) and introducing perturbations in the feature space (such as random spatial dropout, activation dropout, or noise injection). This process encourages the model to learn robust and generalizable feature representations.
Finally, the third stage introduces a **pseudo-labeling strategy**. After the first two stages, the model is already quite capable. In this stage, the model generates its own predictions, or “pseudo labels,” for the unlabeled data. If the model’s confidence in a predicted class for a voxel is above a certain threshold, that prediction is treated as a ground truth label and used for further supervised training. This method helps to further optimize the model weights by leveraging the model’s own confident predictions on unlabeled data, albeit with a reduced loss weighting to mitigate the impact of potential errors.
Also Read:
- CoUn: A Smarter Way for AI Models to Forget Specific Data
- RoSe: A Robust Self-supervised Approach to Stereo Matching in Challenging Weather
Outstanding Performance in Challenges
The U-Mamba2-SSL framework has demonstrated remarkable performance. It achieved an average score of 0.872 and a Dice Similarity Coefficient (DSC) of 0.969 on the validation dataset, showcasing the superior capabilities of this approach. Notably, the framework significantly outperformed other alternatives, securing the first place with an average score of 0.928 in the STSR 2025 Task 1 Challenge validation set. The pre-training stage alone led to a substantial leap in Identification Accuracy (IA), which was further improved by incorporating consistency regularization and pseudo-labeling.
While highly successful, the researchers acknowledge limitations and areas for future work, such as designing specific data processing techniques for different types of CBCTs (full vs. limited field of view) and exploring methods to focus computational resources on regions of interest within the scans.
This multi-stage semi-supervised learning framework represents a significant advancement in automated tooth and pulp segmentation, offering a powerful tool for enhancing precision and efficiency in dental clinical applications. For more detailed information, the full research paper can be accessed here.
The code for U-Mamba2-SSL is publicly available for researchers and developers at https://github.com/zhiqin1998/UMamba2.
