Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Consisaug: A Consistency-based Augmentation for Polyp Detection in Endoscopy Image Analysis

Ziyu Zhou, Wenyuan Shen, Chang Liu

TL;DR

The paper addresses the limited training data and variability in polyp appearance for endoscopy-based detection. It introduces Consisaug, a consistency-based augmentation that uses a Student-Teacher detector and flipping-based augmentations to enforce localization and classification agreement. The main contributions are a localization consistency loss and a Jensen-Shannon-based classification loss, validated on five public polyp datasets and three backbones, with cross-domain transfer demonstrated. The results show robust improvements in recall and mAP50, suggesting practical benefits for more reliable polyp detection in diverse clinical conditions.

Abstract

Colorectal cancer (CRC), which frequently originates from initially benign polyps, remains a significant contributor to global cancer-related mortality. Early and accurate detection of these polyps via colonoscopy is crucial for CRC prevention. However, traditional colonoscopy methods depend heavily on the operator's experience, leading to suboptimal polyp detection rates. Besides, the public database are limited in polyp size and shape diversity. To enhance the available data for polyp detection, we introduce Consisaug, an innovative and effective methodology to augment data that leverages deep learning. We utilize the constraint that when the image is flipped the class label should be equal and the bonding boxes should be consistent. We implement our Consisaug on five public polyp datasets and at three backbones, and the results show the effectiveness of our method.

Consisaug: A Consistency-based Augmentation for Polyp Detection in Endoscopy Image Analysis

TL;DR

The paper addresses the limited training data and variability in polyp appearance for endoscopy-based detection. It introduces Consisaug, a consistency-based augmentation that uses a Student-Teacher detector and flipping-based augmentations to enforce localization and classification agreement. The main contributions are a localization consistency loss and a Jensen-Shannon-based classification loss, validated on five public polyp datasets and three backbones, with cross-domain transfer demonstrated. The results show robust improvements in recall and mAP50, suggesting practical benefits for more reliable polyp detection in diverse clinical conditions.

Abstract

Colorectal cancer (CRC), which frequently originates from initially benign polyps, remains a significant contributor to global cancer-related mortality. Early and accurate detection of these polyps via colonoscopy is crucial for CRC prevention. However, traditional colonoscopy methods depend heavily on the operator's experience, leading to suboptimal polyp detection rates. Besides, the public database are limited in polyp size and shape diversity. To enhance the available data for polyp detection, we introduce Consisaug, an innovative and effective methodology to augment data that leverages deep learning. We utilize the constraint that when the image is flipped the class label should be equal and the bonding boxes should be consistent. We implement our Consisaug on five public polyp datasets and at three backbones, and the results show the effectiveness of our method.
Paper Structure (10 sections, 8 equations, 2 figures, 5 tables)

This paper contains 10 sections, 8 equations, 2 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Method
  3. The Student-Teacher Model
  4. Consistency loss for localization
  5. Consistency loss for classification
  6. Experiments and Results
  7. Implementation Details
  8. Results
  9. Ablation Study
  10. Conclusion

Figures (2)

  • Figure 1: Overall structure of our proposed method.
  • Figure 2: There are three columns for each image set. The first column is the image with ground truth, the second column shows the detection results of vanilla model and the third column is the results of our Consisaug method. The qualitative results prove that our Consisaug can (a) detect small targets, (b) detect targets in motion blur and reflections images, (c) detect targets between colon folds, (d) reduce false positive samples. And in (e) there will also be some failure cases for the hard detecting polyps.