BoxGuide: Exploring Weakly-Supervised Instance Segmentation for Acne Lesions Inspired by Reinforcement Learning

Abstract

Neural network-based instance segmentation methods are essential for the automated diagnosis of patients with acne. However, a lack of pixel-wise annotations, which are time-consuming and labor-intensive to generate, severely impedes the development of segmentation algorithms for acne lesions. To tackle this issue, we propose a novel, task-specific, box-supervised neural network called BoxGuide, the first weakly-supervised instance segmentation approach for acne lesions. By exploiting box annotations, BoxGuide can accurately generate instance masks. Furthermore, two simple yet effective algorithms are utilized in BoxGuide to generate high-quality pseudo labels: Dynamic Ellipse-like Pseudo Label (DEPL) is developed to produce ellipse-like pseudo mask labels with dynamic boundaries, and Center-map Soft Mask (CSM) is constructed to ensure the effectiveness of the generated pseudo mask labels for acne lesions. Utilizing the DEPL and CSM algorithms, the proposed BoxGuide can alleviate the impact of ambiguous boundaries of acne lesions, which greatly limit the performance of conventional general-purpose approaches. Comprehensive experiments on public benchmarks show that BoxGuide greatly surpasses previous approaches. The DEPL algorithm is inspired by the exploration–exploitation concept in reinforcement learning, introducing dynamic boundary perturbation to enhance pseudo labels. BoxGuide only requires box annotations, which significantly reduces the labeling burden on dermatologists.