Low-Contrast-Enhanced Contrastive Learning for Semi-Supervised Endoscopic Image Segmentation

TL;DR

This work tackles endoscopic image segmentation under limited annotations and low-contrast conditions by introducing LoCo, a semi-supervised framework built on a mean-teacher backbone. It leverages low-contrast-enhanced contrastive learning (LCC) with two strategies, inter-class contrast enhancement (ICE) and boundary contrast enhancement (BCE), to learn discriminative features for malignant, benign, and normal tissues. A confidence-based dynamic filter (CDF) optimizes pseudo-label use, especially for minority classes, and a class-aware contrastive loss guided by labeled data strengthens representation. Extensive experiments on a proprietary laryngeal dataset and two public polyp datasets show state-of-the-art performance, with notable gains on minority classes and robust performance under low-label regimes. Together, these contributions advance reliable, data-efficient endoscopic segmentation with practical clinical impact.

Abstract

The segmentation of endoscopic images plays a vital role in computer-aided diagnosis and treatment. The advancements in deep learning have led to the employment of numerous models for endoscopic tumor segmentation, achieving promising segmentation performance. Despite recent advancements, precise segmentation remains challenging due to limited annotations and the issue of low contrast. To address these issues, we propose a novel semi-supervised segmentation framework termed LoCo via low-contrast-enhanced contrastive learning (LCC). This innovative approach effectively harnesses the vast amounts of unlabeled data available for endoscopic image segmentation, improving both accuracy and robustness in the segmentation process. Specifically, LCC incorporates two advanced strategies to enhance the distinctiveness of low-contrast pixels: inter-class contrast enhancement (ICE) and boundary contrast enhancement (BCE), enabling models to segment low-contrast pixels among malignant tumors, benign tumors, and normal tissues. Additionally, a confidence-based dynamic filter (CDF) is designed for pseudo-label selection, enhancing the utilization of generated pseudo-labels for unlabeled data with a specific focus on minority classes. Extensive experiments conducted on two public datasets, as well as a large proprietary dataset collected over three years, demonstrate that LoCo achieves state-of-the-art results, significantly outperforming previous methods. The source code of LoCo is available at the URL of \href{https://github.com/AnoK3111/LoCo}{https://github.com/AnoK3111/LoCo}.

Figures (8)

• Figure 1: Illustration of low-contrast pixels in endoscopic images. Rows $\mathrm{I}$ and $\mathrm{II}$ illustrate the low boundary contrast, which refers to the low contrast between tumor and normal tissues. Rows $\mathrm{III}$ and $\mathrm{IV}$ illustrate the low inter-class contrast, which refers to the low contrast between benign and malignant tumors.
• Figure 2: The overall architecture of LoCo. It is designed based on the mean-teacher framework, featuring two branches: a student network and a teacher network. The teacher network generates reliable pseudo-labels using a confidence-based dynamic filter (CDF). These pseudo-labels, along with labeled images, are combined to supervise the learning of the student network. Additionally, low-contrast-enhanced contrastive learning (LCC) is employed to improve segmentation performance in low-contrast images through inter-class contrast enhancement (ICE) and boundary contrast enhancement (BCE).
• Figure 3: Visualization of the effectiveness of ICE. It is evident that LoCo, when combined with ICE, demonstrates superior inter-class feature similarity compared to the baseline method (i.e., M1).
• Figure 4: Visualization of boundary captured by BCE: (a) Ground truth, (b) Predicted boundary, and (c) Boundary feature similarity.
• Figure 5: Visualization of boundary feature similarity at different epochs: (a) Epoch 1, (b) Epoch 5, (c) Epoch 10, (d) The best epoch, and (e) The best epoch of baseline (M1).