TL;DR
This paper introduces a novel single-stage method for medical image segmentation that jointly trains a segmentation network with a denoising autoencoder (DAE). By learning global anatomical structures through the DAE and constraining the segmentation network with this learned information, the method achieves improved segmentation accuracy and significantly enhances robustness to domain shifts.
What Problem Does This Paper Address?
Medical image segmentation, crucial for tasks like diagnosis and treatment planning, often struggles with producing anatomically implausible results. While deep learning models like U-Net excel at pixel-level accuracy, they can sometimes overlook the broader anatomical context. Existing methods that incorporate anatomical priors often rely on pre-trained autoencoders in a multi-stage process, which can be inefficient and may not fully leverage the learned structural information. This paper addresses the need for a more integrated and effective approach to capture global anatomical structures, leading to more accurate and robust medical image segmentation, particularly in the face of domain variations (e.g., different imaging equipment or patient populations).
Methods
The core of the proposed method is a single-stage joint training framework that combines a segmentation network (specifically, a U-Net architecture) with a denoising convolutional autoencoder (DAE).
- Denoising Autoencoder (DAE): The DAE is trained on ground-truth segmentation labels (binary masks) with added noise (salt-and-pepper noise). Its purpose is to learn a compressed, low-dimensional representation of the anatomical structures present in these labels. The encoder of the DAE learns to map the noisy input to this "anatomical structure space."
- Segmentation Network: This network takes the medical image as input and produces a segmentation output. It consists of an encoder and a decoder.
- Joint Training and Constraint: The key innovation lies in how these two components are linked. The encoder of the segmentation network is constrained to produce feature representations that are similar to the latent space representations learned by the DAE encoder. This is achieved through an embedding loss function (specifically, Mean Squared Error was found to be effective).
- Gradient Flow Control: Crucially, the gradient from this embedding loss is directed only to the segmentation network's encoder, not to the DAE. This prevents the DAE from being negatively impacted by the segmentation task's learning process, allowing it to more effectively learn the underlying anatomical structure information. The DAE is then discarded during inference, acting solely as a training aid.
- Total Loss: The overall loss function is a combination of the segmentation loss (e.g., binary cross-entropy), the DAE reconstruction loss, and the embedding loss, weighted by a parameter $\lambda$.
Datasets
The proposed method was evaluated on two distinct medical image segmentation tasks using publicly available datasets:
- Lung Segmentation:
- JSRT (Japanese Society of Radiological Technology): 247 posterior-anterior (PA) CXR images, with 154 containing nodules and 93 normal.
- MC (Montgomery County): 138 PA CXR images, with 80 normal and 58 from tuberculosis patients.
- Spinal Cord Segmentation:
- Spinal Cord Gray Matter Challenge dataset: 3D MRI images (cut cross-sectionally for 2D processing) from four sites (University College London, Polytechnique Montreal, University of Zurich, Vanderbilt University). Images from one site were reserved for testing due to limited training data.
Key Insights and Ideas
- Single-Stage Integration is Key: Unlike prior work that used pre-trained autoencoders, jointly training the segmentation network and DAE allows for a more dynamic and effective transfer of anatomical knowledge. This avoids the limitations of static, pre-learned priors.
- Anatomical Structure Space: The DAE effectively learns a compressed representation of anatomical structures, creating an "anatomical structure space." By forcing the segmentation network's encoder to map input images into this space, the segmentation process becomes inherently guided by global anatomical context.
- Gradient Control for Robust DAE Learning: Directing the embedding loss gradient only to the segmentation encoder is a clever design choice. It ensures the DAE can focus on learning robust anatomical representations without being perturbed by the segmentation task's specific errors.
- Domain Robustness Through Structure Learning: The paper strongly suggests that learning global anatomical structures is a powerful mechanism for improving robustness to domain shifts. By understanding the underlying anatomy, the model becomes less sensitive to variations in imaging equipment, protocols, or patient populations.
Key Results
- Improved Segmentation Performance: The proposed method consistently outperformed baseline U-Net, ACNN, and SRM on both lung and spinal cord segmentation tasks across various metrics (IOU, DSC, ACD, ASD).
- For lung segmentation on JSRT, the proposed method achieved an ASD of 0.691, outperforming SRM by 5.6%.
- For spinal cord segmentation on site2, the proposed method achieved an ASD of 0.508, 2.6% lower than the second-best model ACNN.
- Enhanced Domain Robustness: The method demonstrated significant improvements in domain generalization.
- In lung segmentation, when trained on JSRT and tested on MC, the proposed method achieved an average IOU of 0.924, outperforming ACNN by 2.0%.
- For spinal cord segmentation, averaged across three domain shift settings, the proposed method achieved an average ACD of 0.667, a 9.7% improvement over the baseline U-Net.
- Qualitative Improvements: Visualizations showed that the proposed method produced more accurate segmentation boundaries and fewer false positives, especially in challenging cases and across different domains. The reconstruction results from the combined DAE and segmentation decoder (fSEG → gDAE) also indicated that the learned features were anatomically informative.
Limitations and Future Work
- DAE Discarded at Inference: The DAE is used solely during training and is not part of the inference pipeline. While this simplifies deployment, it means the learned anatomical prior is implicitly encoded within the segmentation encoder's weights.
- Hyperparameter Tuning: The weight of the embedding loss ($\lambda$) needs to be determined through validation, which is a standard but necessary step.
- Future Work: The authors suggest that this approach could be particularly beneficial for other medical applications requiring precise anatomical understanding, such as cranial implant design or tooth segmentation. The findings highlight the generalizability of learning global anatomical structures for robust segmentation across diverse medical imaging scenarios.
Summary updated