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

TL;DR

Monobox tackles the practical challenge of box-supervised polyp segmentation under non-tight box annotations by introducing a monotonicity constraint that governs unconfident regions around box edges. The framework combines a proxy-map optimization with a dynamic label-correction mechanism, enabling robust learning without precise box tightness and progressively improving annotation quality. Empirical results show consistent Dice gains over state-of-the-art anti-noise MIL-based methods on both synthetic and real noisy datasets, and demonstrate strong generalization to other MIL-based BSS tasks such as COCO BoxInst. The approach reduces annotation burden and enhances clinical applicability by tolerating imperfect boxes while maintaining segmentation accuracy.

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.

Figures (7)

• Figure 1: Examples of non-tight box annotations produced by endoscopists in the real annotation process. The red dashed lines indicate regions where the annotation is too wide, and the yellow dashed lines indicate regions where the annotation is too narrow.
• Figure 2: In cases (1, 2, 3, 4) of four typical noisy box annotations (green line), our proposed monotonicity constraint (MC) provides correct constrains for four sampled bags (black dashed line) from the unconfident region (the region between two yellow lines), but the traditional MIL lead to incorrect constraints. For brevity, we only visualize the local region of the upper boundary of the box annotation.
• Figure 3: (a) Schema of generating GT box-filled mask (assigning 1 within box annotations and 0 outside box annotations), confident region (as Eq. \ref{['eq4']}) and unconfident regions (as Eq. \ref{['eq3']}). (b) Overview of our proposed MonoBox. For the proxy map decoupled from the segmented map, we first define the confident and unconfident regions, and then adopt the consistency constraint and monotonicity constraint on them, respectively. Moreover, we utilize a strategy of label correction strategy to dynamically improve the tightness of box annotations
• Figure 4: Visualization results of the different anti-noise methods using IBoxCLA as the model on the Real noisy dataset and the Synthetic noisy dataset.
• Figure 5: Visualization results of the baseline and three variants. (a)-(d) correspond to the 1st to the 4th rows in Table \ref{['table3']}, respectively. Yellow arrows show the predicted small thorn-like incorrect regions.