Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

MixPolyp: Integrating Mask, Box and Scribble Supervision for Enhanced Polyp Segmentation

Yiwen Hu, Jun Wei, Yuncheng Jiang, Haoyang Li, Shuguang Cui, Zhen Li, Song Wu

TL;DR

The mixed supervised polyp segmentation paradigm (MixPolyp) combines diverse annotation types (mask, box, and scribble) within a single model, thereby expanding the range of available data and reducing labeling costs.

Abstract

Limited by the expensive labeling, polyp segmentation models are plagued by data shortages. To tackle this, we propose the mixed supervised polyp segmentation paradigm (MixPolyp). Unlike traditional models relying on a single type of annotation, MixPolyp combines diverse annotation types (mask, box, and scribble) within a single model, thereby expanding the range of available data and reducing labeling costs. To achieve this, MixPolyp introduces three novel supervision losses to handle various annotations: Subspace Projection loss (L_SP), Binary Minimum Entropy loss (L_BME), and Linear Regularization loss (L_LR). For box annotations, L_SP eliminates shape inconsistencies between the prediction and the supervision. For scribble annotations, L_BME provides supervision for unlabeled pixels through minimum entropy constraint, thereby alleviating supervision sparsity. Furthermore, L_LR provides dense supervision by enforcing consistency among the predictions, thus reducing the non-uniqueness. These losses are independent of the model structure, making them generally applicable. They are used only during training, adding no computational cost during inference. Extensive experiments on five datasets demonstrate MixPolyp's effectiveness.

MixPolyp: Integrating Mask, Box and Scribble Supervision for Enhanced Polyp Segmentation

TL;DR

The mixed supervised polyp segmentation paradigm (MixPolyp) combines diverse annotation types (mask, box, and scribble) within a single model, thereby expanding the range of available data and reducing labeling costs.

Abstract

Limited by the expensive labeling, polyp segmentation models are plagued by data shortages. To tackle this, we propose the mixed supervised polyp segmentation paradigm (MixPolyp). Unlike traditional models relying on a single type of annotation, MixPolyp combines diverse annotation types (mask, box, and scribble) within a single model, thereby expanding the range of available data and reducing labeling costs. To achieve this, MixPolyp introduces three novel supervision losses to handle various annotations: Subspace Projection loss (L_SP), Binary Minimum Entropy loss (L_BME), and Linear Regularization loss (L_LR). For box annotations, L_SP eliminates shape inconsistencies between the prediction and the supervision. For scribble annotations, L_BME provides supervision for unlabeled pixels through minimum entropy constraint, thereby alleviating supervision sparsity. Furthermore, L_LR provides dense supervision by enforcing consistency among the predictions, thus reducing the non-uniqueness. These losses are independent of the model structure, making them generally applicable. They are used only during training, adding no computational cost during inference. Extensive experiments on five datasets demonstrate MixPolyp's effectiveness.
Paper Structure (11 sections, 5 equations, 3 figures, 2 tables)

This paper contains 11 sections, 5 equations, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Method
  3. Box Supervision with Subspace Projection Loss
  4. Scribble Supervision with Binary Minimum Entropy Loss
  5. Linear Regularization Loss
  6. Overall Training Loss Function
  7. Experiment
  8. Dataset and Implementation Details
  9. Performance Comparison
  10. Conclusion
  11. Acknowledgement

Figures (3)

  • Figure 1: Illustration of our MixPolyp framework. It consists of three learning branches: (1) Full supervision branch for pixel-level annotation data, (2) Box supervision branch with Subspace Projection loss ($\mathcal{L_{SP}}$), and (3) Scribble supervision branch with Binary Minimum Entropy loss ($\mathcal{L_{BME}}$). In addition, in both box supervision and scribble supervision branches, we introduce Linear Regularization loss ($\mathcal{L_{LR}}$) to constrain the consistency between predictions.
  • Figure 2: Subspace Projection Loss, which first projects the predicted mask and the box annotations into 1D vectors and then calculates the supervision loss between these vectors.
  • Figure 3: Loss curve comparison between Binary Cross Entropy Loss and Binary Minimum Entropy Loss.