MonoBox: Tightness-free Box-supervised Polyp Segmentation using Monotonicity Constraint

Qiang Hu, Zhenyu Yi, Ying Zhou, Ting Li, Fan Huang, Mei Liu, Qiang Li, Zhiwei Wang

Abstract

We propose MonoBox, an innovative box-supervised segmentation method constrained by monotonicity to liberate its training from the user-unfriendly box-tightness assumption. In contrast to conventional box-supervised segmentation, where the box edges must precisely touch the target boundaries, MonoBox leverages imprecisely-annotated boxes to achieve robust pixel-wise segmentation. The 'linchpin' is that, within the noisy zones around box edges, MonoBox discards the traditional misguiding multiple-instance learning loss, and instead optimizes a carefully-designed objective, termed monotonicity constraint. Along directions transitioning from the foreground to background, this new constraint steers responses to adhere to a trend of monotonically decreasing values. Consequently, the originally unreliable learning within the noisy zones is transformed into a correct and effective monotonicity optimization. Moreover, an adaptive label correction is introduced, enabling MonoBox to enhance the tightness of box annotations using predicted masks from the previous epoch and dynamically shrink the noisy zones as training progresses. Experiments on both public synthetic and in-house real noisy datasets demonstrate that MonoBox exceeds other anti-noise state-of-the-arts by improving Dice by at least 5.5% and 3.3%, respectively.

Introduction

In medical imaging, training deep learning models for segmentation typically requires dense, pixel-perfect masks. However, generating these masks is labor-intensive and requires significant clinical expertise. Box-supervised Segmentation (BSS) has emerged as a popular alternative, where annotators provide simple bounding boxes around targets like polyps.

Despite its efficiency, BSS traditionally relies on a strict "tightness assumption": the box edges must precisely touch the object's boundaries. In clinical practice, particularly with polyps that exhibit blurry edges or low contrast against normal tissue, endoscopists often produce "non-tight" boxes. Traditional models based on Multiple Instance Learning (MIL) often fail under these conditions, as they attempt to learn from inaccurate binary labels, leading to poor segmentation performance and artifacts.

Methodology

MonoBox addresses box non-tightness as a noise-learning problem through three primary innovations designed to move beyond rigid binary supervision.

Proxy Map Optimization

Instead of optimizing the segmentation map directly, the model generates a "proxy map" that decouples shape information. This architectural choice facilitates the application of constraints across entire rows and columns of the image, allowing the model to learn structural trends rather than just isolated pixel values.

The Monotonicity Constraint (MC)

The core of the framework is the Monotonicity Constraint (MC). The authors define "unconfident regions" or noisy zones around the box edges where the exact boundary is uncertain. Rather than forcing the model to match a potentially incorrect binary label (0 or 1) in these zones, MC enforces a relative trend: the model’s response must monotonically decrease as it moves from the foreground center toward the background.

Comparison of Multiple Instance Learning (MIL) and Monotonicity Constraint (MC). MIL fails (red crosses) when box edges do not align perfectly with object boundaries, whereas MC maintains a valid trend (green checks).

This shift from absolute classification to relative trend-following makes the model robust to human error. By optimizing for a decreasing gradient of confidence, the model effectively "prunes" the common "thorn" artifacts—tiny splinters of predicted mask that reach out to touch the edges of oversized, loose boxes.

Adaptive Label Correction (LC)

As training progresses and the model's predictions become more reliable, MonoBox employs Adaptive Label Correction (LC). The framework uses predicted masks from previous epochs to "shrink-wrap" the original noisy boxes. This dynamically tightens the annotations and reduces the size of the unconfident regions, allowing the model to transition from coarse trend-following to more precise boundary refinement.

Experimental Evaluation

Datasets and Setup

The researchers validated MonoBox across diverse scenarios:

  1. Public Synthetic Noisy Datasets: Utilizing five standard polyp datasets (ClinicDB, Kvasir-SEG, ColonDB, EndoScene, and ETIS), the team simulated "sloppy" annotations by randomly shifting and scaling ground-truth boxes.
  2. In-house Real Noisy Dataset: A large-scale collection of 18,656 colonoscopy images from a local hospital featuring real-world, non-tight box annotations provided by practicing endoscopists.
Representative examples of box-supervised annotations in colonoscopy images, highlighting the challenges of vague boundaries.

Performance and Robustness

MonoBox demonstrated significant improvements over existing state-of-the-art anti-noise methods. On synthetic noisy data, it improved the Dice score by at least 5.5%, and on real-world clinical data, it achieved a 3.3% improvement. Notably, as the noise level (the degree of box looseness) increased, MonoBox's performance remained stable, whereas traditional MIL-based methods saw sharp declines in accuracy.

Discussion and Conclusion

A key advantage of MonoBox is its "plug-and-play" nature; it is a constraint and loss strategy that can be integrated into existing BSS backbones like WeakPolyp or IBoxCLA. Qualitative results indicate that MonoBox produces smoother, more anatomically plausible boundaries compared to the jagged outputs of previous methods.

While highly effective, the authors note that the current iteration applies a uniform "unconfident scale" to all images in a batch. Future work will focus on adaptive unconfident scales, allowing the model to estimate the reliability of each individual box based on local image characteristics like contrast and blur. In summary, MonoBox represents a significant step toward making weakly-supervised medical image segmentation practical for real-world clinical environments where perfect annotations are unavailable.

Summary updated

Abstract

We propose MonoBox, an innovative box-supervised segmentation method constrained by monotonicity to liberate its training from the user-unfriendly box-tightness assumption. In contrast to conventional box-supervised segmentation, where the box edges must precisely touch the target boundaries, MonoBox leverages imprecisely-annotated boxes to achieve robust pixel-wise segmentation. The 'linchpin' is that, within the noisy zones around box edges, MonoBox discards the traditional misguiding multiple-instance learning loss, and instead optimizes a carefully-designed objective, termed monotonicity constraint. Along directions transitioning from the foreground to background, this new constraint steers responses to adhere to a trend of monotonically decreasing values. Consequently, the originally unreliable learning within the noisy zones is transformed into a correct and effective monotonicity optimization. Moreover, an adaptive label correction is introduced, enabling MonoBox to enhance the tightness of box annotations using predicted masks from the previous epoch and dynamically shrink the noisy zones as training progresses. We verify MonoBox in the box-supervised segmentation task of polyps, where satisfying box-tightness is challenging due to the vague boundaries between the polyp and normal tissues. Experiments on both public synthetic and in-house real noisy datasets demonstrate that MonoBox exceeds other anti-noise state-of-the-arts by improving Dice by at least 5.5% and 3.3%, respectively. Codes are at https://github.com/Huster-Hq/MonoBox.
BibTeX
@inproceedings{hu2025monobox,
  author = {Qiang Hu and Zhenyu Yi and Ying Zhou and Ting Li and Fan Huang and Mei Liu and Qiang Li and Zhiwei Wang},
  title = {{MonoBox: Tightness-free Box-supervised Polyp Segmentation using Monotonicity Constraint}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i4.32371},
}
Figure