ReFit: A Framework for Refinement of Weakly Supervised Semantic Segmentation using Object Border Fitting for Medical Images

  • 2023-09-20 12:06:23
  • Bharath Srinivas Prabakaran, Erik Ostrowski, Muhammad Shafique
  • 0

Abstract

Weakly Supervised Semantic Segmentation (WSSS) relying only on image-levelsupervision is a promising approach to deal with the need for Segmentationnetworks, especially for generating a large number of pixel-wise masks in agiven dataset. However, most state-of-the-art image-level WSSS techniques lackan understanding of the geometric features embedded in the images since thenetwork cannot derive any object boundary information from just image-levellabels. We define a boundary here as the line separating an object and itsbackground, or two different objects. To address this drawback, we areproposing our novel ReFit framework, which deploys state-of-the-art classactivation maps combined with various post-processing techniques in order toachieve fine-grained higher-accuracy segmentation masks. To achieve this, weinvestigate a state-of-the-art unsupervised segmentation network that can beused to construct a boundary map, which enables ReFit to predict objectlocations with sharper boundaries. By applying our method to WSSS predictions,we achieved up to 10% improvement over the current state-of-the-art WSSSmethods for medical imaging. The framework is open-source, to ensure that ourresults are reproducible, and accessible online athttps://github.com/bharathprabakaran/ReFit.

 

Quick Read (beta)

loading the full paper ...