Abstract
The predictive performance of segmentation highly depends on the quality of annotations. However, in the case of medical images, it is difficult to annotate precise regions of disease since they have a variety of shapes and ambiguous boundaries. Improving segmentation performance of models trained with rough annotations will make it easier to create training data. This paper proposes a novel method to improve auto tumor segmentation from computed tomography (CT) images using noisy annotations. The proposed method uses pre-training to detect rough tumor regions and fine-tuning to refine the roughly detected regions. In order to refine roughly detected regions, Active Contours Dice (ACD) Loss is introduced in fine-tuning. In our experiments using simulated rough annotations, the proposed method improves segmentation performance.
Introduction
In medical imaging, particularly for CT scans, tumors and lesions often exhibit ambiguous or "unclear" boundaries. For human experts, tracing these boundaries perfectly is a Herculean task, leading to datasets that frequently contain "rough" annotations—quick circles or loose polygons that include significant boundary errors. Standard segmentation models, such as the 3D U-Net, are highly sensitive to these errors; when trained on sloppy data, they tend to produce sloppy results. This research addresses a critical gap: building a system that accepts low-quality, noisy annotations as input but outputs high-quality, precise segmentations.
Proposed Methodology: Two-Stage Training
The authors propose a robust two-stage training strategy utilizing a 3D U-Net architecture to bridge the gap between noisy labels and precise boundaries.
- Pre-training (Baseline): The model is first trained on rough annotations using standard loss functions like Dice loss. This stage allows the model to learn the general location and approximate shape of the target disease.
- Fine-tuning with ACD Loss: This is the core innovation. The authors introduce the Active Contours Dice (ACD) Loss. Inspired by traditional "snakes" or active contour models, this loss function calculates the difference between the model's current prediction and a version of that prediction that has been "pushed" toward the nearest sharp edges in the actual CT image using Gaussian derivatives and Sobel filters.
By combining ACD Loss with a standard region-based loss during fine-tuning, the model is incentivized to remain within the annotated area while "snapping" its boundaries to the actual physical gradients found in the raw image data.
Experimental Setup and Datasets
The study utilizes the LiTS (Liver Tumor Segmentation) dataset, a well-known benchmark in medical imaging. To evaluate the robustness of the method, the authors created a "rough" version of the dataset by programmatically distorting expert ground truth. This involved applying random rotations and scaling (between 1.8x to 2.2x the original size) to mimic oversized, loose annotations typical of rushed human labeling.
Key Insights and Results
The "aha moment" of this research is the realization that while human labels might be noisy, the raw image data—specifically the gradients and edges—serves as a source of objective truth. The ACD loss acts as a bridge, allowing raw image features to "correct" the human annotations. The authors found that starting with the complex ACD loss from scratch is unstable; the two-stage approach is crucial for convergence.
The proposed method showed significant improvements over both the baseline and previous state-of-the-art methods for handling noisy labels:
- Dice Score: The proposed method achieved a score of 0.430, compared to the baseline's 0.376 and a previous "Noise-Robust Loss" method which scored 0.322.
- IoU (Intersection over Union): Improved to 0.322 from the baseline's 0.265.
Qualitative results demonstrate the model's ability to successfully ignore the "extra" area provided in rough training annotations, effectively shrinking the predicted mask to tightly hug the actual tumor boundaries.
Limitations and Future Work
Despite its success, the method relies on image gradients. If a tumor boundary is so faint that there is almost no contrast change in the CT scan, the ACD loss has no "edge" to snap to. Additionally, the use of 3D convolutions and derivative calculations increases computational overhead. Future work may involve automating the balance between region loss and ACD loss and validating the method on a wider variety of naturally noisy datasets where human errors are more diverse.
Summary updated