Abstract
Computed Tomography (CT) plays an important role in monitoring radiation-induced Pulmonary Fibrosis (PF), where accurate segmentation of the PF lesions is highly desired for diagnosis and treatment follow-up. However, the task is challenged by ambiguous boundary, irregular shape, various position and size of the lesions, as well as the difficulty in acquiring a large set of annotated volumetric images for training. To overcome these problems, we propose a novel convolutional neural network called PF-Net and incorporate it into a semi-supervised learning framework based on Iterative Confidence-based Refinement And Weighting of pseudo Labels (I-CRAWL). Our PF-Net combines 2D and 3D convolutions to deal with CT volumes with large inter-slice spacing, and uses multi-scale guided dense attention to segment complex PF lesions. For semi-supervised learning, our I-CRAWL employs pixel-level uncertainty-based confidence-aware refinement to improve the accuracy of pseudo labels of unannotated images, and uses image-level uncertainty for confidence-based image weighting to suppress low-quality pseudo labels in an iterative training process. Extensive experiments with CT scans of Rhesus Macaques with radiation-induced PF showed that: 1) PF-Net achieved higher segmentation accuracy than existing 2D, 3D and 2.5D neural networks, and 2) I-CRAWL outperformed state-of-the-art semi-supervised learning methods for the PF lesion segmentation task.
Introduction
Radiation-induced Pulmonary Fibrosis (PF) is a common side effect for lung cancer patients undergoing radiotherapy. Monitoring these lesions via CT scans is vital for treatment, but manual segmentation is a significant challenge for radiologists. PF lesions are irregular in shape, vary wildly in size, and often have "fuzzy" boundaries that blend into surrounding vessels or pneumonia.
Technically, two major hurdles exist in automated segmentation:
- Anisotropic Resolution: Medical CT scans often have high resolution within a slice but large gaps between slices (inter-slice spacing). Standard 3D CNNs assume isotropic pixels and struggle when data is anisotropic.
- The Annotation Bottleneck: Training deep models requires massive amounts of labeled data. In medical imaging, obtaining expert 3D labels for hundreds of scans is often unfeasible.
Methods
PF-Net Architecture
Instead of using a pure 2D or 3D network, PF-Net utilizes a 2.5D approach. It uses 2D convolutions at high-resolution scales to capture fine in-plane details and switches to 3D convolutions at lower resolutions to capture spatial context between slices. This hybrid design balances the physical receptive field, preventing bias from large inter-slice gaps.
It also features Multi-Scale Guided Dense Attention (MSDA). Unlike standard attention mechanisms, PF-Net propagates attention maps from all coarser scales down to finer scales. This dense connection helps the network maintain focus on the lesion across different sizes and resolutions.
I-CRAWL Training Framework
To leverage unlabeled data, the authors propose I-CRAWL, an iterative semi-supervised framework:
- Uncertainty Estimation: The model uses "Monte Carlo Dropout" to identify regions where it is unsure of its predictions.
- Confidence-Aware Refinement (CAR): In high-uncertainty areas, the model uses a Conditional Random Field (CRF) to smooth labels based on surrounding pixel context.
- Image Weighting: If a generated pseudo-label is globally uncertain, the framework down-weights that image during training to avoid "poisoning" the model with inaccurate data.
Experimental Results
The study evaluated the framework using thoracic CT scans from 41 male rhesus macaques. In the primary semi-supervised setup, only 20% of the scans were labeled.
Key Performance Metrics
- Superior Accuracy: PF-Net achieved a Dice score of 70.36%, significantly outperforming standard 2D U-Nets (59.3%) and 3D V-Nets (60.4%).
- Efficiency with Limited Data: Using I-CRAWL with only 20% labeled data, the model reached a Dice score of 73.04%, nearly matching the performance of full supervision (74.58%).
- Boundary Robustness: The model showed a massive improvement in the 95th percentile Hausdorff Distance (HD95), indicating much cleaner and more accurate predicted boundaries.
Conclusion
PF-Net and the I-CRAWL framework demonstrate that combining hybrid architectures with uncertainty-aware semi-supervised learning can effectively overcome the challenges of medical image segmentation. While the current study used an animal model, the methodology is highly applicable to human clinical datasets for pathologies like COVID-19 or lung tumors. Future work will focus on reducing the computational cost of uncertainty estimation and validating the model in broader clinical settings.
Summary updated