Related Experiment Video
Updated: Jun 15, 2025
04:48
Swin-PSAxialNet: An Efficient Multi-Organ Segmentation Technique
Published on: July 5, 2024
379
Enhanced accuracy with Segmentation of Colorectal Polyp using NanoNetB, and Conditional Random Field Test-Time
Muhammad Sajjad Hussain1, Umer Asgher2,3, Sajid Nisar4
1Department of Computer Science, Sir Syed (CASE) Institute of Technology, Islamabad, Pakistan.
Frontiers in Robotics and AI
|August 26, 2024
Summary
This study introduces an Enhanced Nanonet model for precise, real-time polyp segmentation in colonoscopy images, significantly reducing missed detections. The lightweight model improves accuracy, especially for small polyps, aiding early colorectal cancer prevention.
Area of Science:
- Medical Imaging
- Artificial Intelligence in Medicine
- Gastroenterology
Background:
- Colonoscopy is crucial for early colorectal polyp detection and cancer prevention.
- Current colonoscopy techniques suffer from high polyp miss rates due to challenges like polyp shape, size, and resemblance to surrounding tissues.
- Existing deep learning models for polyp segmentation are often too complex for real-time clinical application, requiring lightweight and efficient solutions.
Purpose of the Study:
- To develop a novel, lightweight, and generalized Enhanced Nanonet model for accurate, real-time polyp segmentation in colonoscopy images.
- To improve upon the Nanonet architecture using NanonetB, incorporating data augmentation, Conditional Random Field (CRF), and Test-Time Augmentation (TTA).
- To address the challenges of real-time segmentation, minimal latency, and seamless integration with endoscopic hardware.
Main Methods:
- Proposed a novel Enhanced Nanonet model, an improvement over Nanonet using NanonetB.
- Implemented data augmentation, Conditional Random Field (CRF), and Test-Time Augmentation (TTA) to enhance model performance.
- Evaluated the model on six diverse, publicly available colonoscopy datasets (Kvasir-SEG, Endotect Challenge 2020, Kvasir-instrument, CVC-ClinicDB, CVC-ColonDB, CVC-300) for generalizability.
Main Results:
- Achieved a mean Intersection over Union (mIoU) of 0.8188 and a Dice coefficient of 0.8060 on the Kvasir-SEG dataset.
- The Enhanced Nanonet model has only 132,049 parameters, requiring minimal computational resources.
- Demonstrated improved performance in detecting smaller and sessile (flat) polyps, which are often missed in standard examinations.
Conclusions:
- The proposed Enhanced Nanonet model offers a lightweight and generalized solution for precise, real-time polyp segmentation in colonoscopy.
- The integration of CRF and TTA significantly enhances segmentation performance across different datasets, validating its generalizability.
- This model has the potential to reduce miss rates in colonoscopy, thereby improving early detection and prevention of colorectal cancer.
