Introduction
Training deep learning models for medical image segmentation typically requires expensive, pixel-level masks. Box-Supervised Segmentation (BSS) has emerged as a popular alternative, utilizing simple bounding boxes that are significantly faster to produce. However, most BSS methods rely on a "tightness assumption," requiring box edges to perfectly touch object boundaries. MonoBox is a novel framework that eliminates this requirement by enforcing a monotonicity constraint, ensuring robust training even when clinicians provide loose or imprecise annotations.
The Problem: Limitations of the Tightness Assumption
In clinical practice, the tightness assumption often fails. Polyps frequently exhibit blurry boundaries and low contrast with surrounding tissue, leading endoscopists to draw boxes that are either too wide or too narrow.
Traditional BSS methods, often based on Multiple Instance Learning (MIL), struggle with these "noisy" edges. When a box is oversized, the model attempts to force the segmentation to touch the incorrect boundaries, frequently producing "thorn-like" artifacts and splinters that degrade the clinical utility of the output.
Methods: Monotonicity Constraint and Label Correction
MonoBox introduces two primary mechanisms to handle non-tight boxes and improve segmentation accuracy:
- Monotonicity Constraint (MC): Instead of forcing the model to match potentially incorrect binary labels in "unconfident regions" near box edges, MonoBox enforces a spatial trend. Using four directional kernels, the MC ensures that the predicted probability of a pixel being a polyp decreases monotonically as one moves from the center of the box toward the background. This provides a reliable gradient for optimization without over-fitting to noisy labels.
- Adaptive Label Correction (LC): As the model's confidence grows during training, it utilizes its own predicted masks to refine the original bounding boxes. Periodically, the framework "tightens" the boxes based on high-probability predictions and dynamically shrinks the unconfident regions, creating a virtuous cycle that gradually cleans the noisy dataset.
Additionally, the framework optimizes a Proxy Map, which decouples shape information from the segmentation map, allowing for more direct optimization of the box-level constraints.
Datasets and Evaluation
The researchers evaluated MonoBox across three distinct scenarios:
- Public Synthetic Noisy Dataset: A combination of five public polyp datasets (including ClinicDB and Kvasir-SEG) where ground-truth boxes were perturbed with Gaussian noise to simulate inaccuracy.
- In-House Real Noisy Dataset: A large-scale dataset of 18,656 colonoscopy images featuring real-world annotations from endoscopists, which naturally contained non-tight boxes.
- COCO Dataset: To demonstrate the framework's generality, tests were also conducted on natural images using the COCO test-dev split.
Key Results and Insights
MonoBox significantly outperformed existing state-of-the-art anti-noise methods:
- Performance Gains: On the real-world noisy dataset, MonoBox improved the Dice score by 3.3% over the previous best method (FSRM). On synthetic noisy data, the improvement reached 5.5%.
- Visual Quality: Qualitative analysis confirmed that MonoBox successfully eliminates the "thorn-like" artifacts common in other box-supervised methods when boxes are over-sized.
- Robustness: The core innovation is the shift from "targets to trends." By exploiting the spatial correlation of noise—recognizing that neighbors closer to the center are more likely to be foreground—the method remains stable even as box tightness decreases.
Limitations and Future Work
Currently, MonoBox utilizes a uniform scale for the "unconfident region" across all samples in a dataset. However, in practice, some boxes are noisier than others. Future research will explore methods to adaptively perceive the appropriate unconfident scale for each individual training sample, potentially using the model's own uncertainty metrics to determine how much of a box edge should be distrusted at any given time.
Summary updated