Source-linked AI summary
Polyp Segmentation in Colonoscopy Images Using Fully Convolutional Network
Mojtaba Akbari, Majid Mohrekesh, Ebrahim Nasr-Esfahani, S. M. Reza Soroushmehr, Nader Karimi, Shadrokh Samavi, Kayvan Najarian
TL;DR
Colonoscopy polyp segmentation is difficult because polyp appearance varies, while early detection matters for colorectal cancer care. The paper proposes a CNN-based pipeline with smart patch selection during training and probability-map post-processing during testing, achieving an 81% Dice score on CVC-ColonDB and outperforming previous methods.
Problem
Colonoscopy polyp segmentation is challenging because polyps vary in shape, color intensity, size, and appearance, motivating computer-aided methods.
Method
The method uses FCN-8S to generate candidate regions, smart patch selection for training, and Otsu thresholding with largest-connected-component selection for post-processing.
Results
81% Dice score was achieved on CVC-ColonDB, outperforming previous colorectal-polyp segmentation methods.
Takeaways & Limitations
The reported results support combining CNN-based segmentation with targeted patch selection and post-processing for colonoscopy polyp segmentation.
Abstract
from arXiv · showhide
Colorectal cancer is a one of the highest causes of cancer-related death, especially in men. Polyps are one of the main causes of colorectal cancer and early diagnosis of polyps by colonoscopy could result in successful treatment. Diagnosis of polyps in colonoscopy videos is a challenging task due to variations in the size and shape of polyps. In this paper we proposed a polyp segmentation method based on convolutional neural network. Performance of the method is enhanced by two strategies. First, we perform a novel image patch selection method in the training phase of the network. Second, in the test phase, we perform an effective post processing on the probability map that is produced by the network. Evaluation of the proposed method using the CVC-ColonDB database shows that our proposed method achieves more accurate results in comparison with previous colonoscopy video-segmentation methods.
I. INTRODUCTION
Colorectal polyps are important targets for early detection, but colonoscopy segmentation is difficult because polyp appearance varies. The paper proposes a CNN-based method enhanced by patch selection and post-processing.
- Colorectal cancer is a major cause of death, and early segmentation and detection can increase survival chances.
- Colonoscopy-based polyp segmentation is challenging because polyps vary in shape and color intensity.
- Computer-aided segmentation is needed because colonoscopy depends on operators, whose mistakes and limited sensitivity can affect analysis.
- Existing polyp-segmentation research includes image-processing methods, feature extraction with classifiers, and CNN-based approaches.
- The proposed method combines cascading CNNs, smart patch selection, adaptive thresholding, and largest-connected-component selection.
II. RELATED WORKS
Prior polyp-segmentation methods span image processing, hand-crafted feature classification, CNNs, FCNs, U-Net, and CNN ensembles. The paper positions smart patch selection as a response to limited medical-image training data.
- Image-processing approaches use region properties and image cues such as valley depth to identify polyp candidates.
- Feature-based methods extract descriptors from image patches and classify them as polyp or non-polyp.
- CNN methods learn features from patches or images and use classifiers or smoothing to produce polyp-segmentation results.
- FCN and U-Net replace fully connected layers with deconvolution-based prediction maps that incorporate information from earlier layers.
- CNN ensembles combine multiple networks to address variation in polyp shapes and intensities.
- The paper identifies limited medical-image databases as a training challenge and presents smart patch selection to address it.
III. PROPOSED METHOD
The proposed segmentation pipeline has two stages: FCN-8S generates probable polyp candidate regions, then thresholding and connected-component selection isolate the segmented region.
- FCN-8S first proposes candidate regions likely to contain polyps.
- Otsu thresholding converts the FCN probability map into a binary image.
- The largest connected component is selected as the polyp region among the candidate regions.
A. FCN-8S Network
FCN creates dense prediction maps through convolution, pooling, deconvolution, and upsampling. The proposed method uses FCN-8S, which combines features from pool3, pool4, and conv7.
- FCN generates dense feature maps through convolution and pooling, then enlarges them with deconvolution and upsampling.
- FCN-32S, FCN-16S, and FCN-8S differ in their upsampling scales and the earlier-layer features used for prediction.
- FCN-8S combines pool3, pool4, and conv7 outputs and upsamples by scale 8 to produce an input-sized prediction map.
- The proposed method uses FCN-8S for colonoscopy polyp-region segmentation and applies Otsu thresholding followed by largest-component selection.
B. Patch Selection and Data Augmentation
Training FCN-8S is constrained by limited medical-image data and ground truth, so the method uses augmentation and intelligent patch selection to improve training coverage.
- Limited medical-image data and ground truth make FCN-8S training an important challenge.
- Image rotation exposes FCN to different polyp structures across images.
- Patch selection samples centers from polyp interiors, borders, and background regions.
IV. EXPERIMENTAL RESULTS
The method was evaluated on physician-annotated CVC-ColonDB images using augmented FCN-8S training, patch selection, and post-processing, with improved segmentation results over prior methods.
- CVC-ColonDB contains 300 physician-annotated polyp images from 15 distinct-study sequences.Each sequence has resolution 500×574 in RGB color space.
- Patch selection and largest-connected-component post-processing were evaluated using Dice scores and False Positive rate Per Frame.The reported comparisons include original images, rotation, and random patch selection.
- The proposed method achieves higher sensitivity with lower FPPF than and.
- The proposed method achieves better quality assessment results than the method of.
V. CONCLUSION
The paper proposes CNN-based polyp segmentation with Otsu thresholding and wise patch selection, achieving an 81% Dice score on CVC-ColonDB and outperforming previous methods.
- The proposed method combines convolutional neural networks, Otsu thresholding, and wise patch selection for polyp segmentation.
- 81% Dice score was achieved on CVC-ColonDB, outperforming previous colorectal-polyp segmentation methods.