Polygonal Approximation Learning for Convex Object Segmentation in Biomedical Images With Bounding Box Supervision

Area of Science:

• Medical Image Analysis
• Computer Vision
• Biomedical Engineering

Background:

• Deep learning for biomedical image segmentation requires extensive, fine-grained annotations, which are costly and time-consuming.
• Existing methods struggle with data dependence, limiting their application in resource-constrained scenarios.
• Bounding-box supervision offers a less labor-intensive alternative but often yields suboptimal segmentation results.

Purpose of the Study:

• To introduce Polygonal Approximation Learning (PAL), a novel approach for convex object instance segmentation using only bounding-box supervision.
• To demonstrate that detection models for convex objects inherently contain segmentation information derivable from bounding boxes.
• To validate PAL's effectiveness and compare its performance against existing methods in biomedical imaging.

Main Methods:

• Proposed Polygonal Approximation Learning (PAL) for instance segmentation with bounding-box supervision.
• Utilized a repeated detection approach with rotated biomedical images to extract segmentation information from detection models.
• Employed a dice loss function incorporating projections of rotated detection results for training the segmentation model.

Main Results:

• PAL significantly outperforms existing box-supervised models like BoxInst for convex object instance segmentation (e.g., nuclei).
• PAL achieves performance comparable to mask-supervised models such as Mask R-CNN and Cascade Mask R-CNN.
• Demonstrated remarkable performance of PAL on non-convex object instance segmentation tasks, including surgical instruments and organs.

Conclusions:

• PAL effectively alleviates the data dependence in deep learning-based biomedical image segmentation by leveraging bounding-box supervision.
• The method provides a cost-effective and efficient alternative to traditional annotation-intensive approaches.
• PAL shows broad applicability across various biomedical imaging tasks involving both convex and non-convex object segmentation.