# Multi-stage convo-enhanced retinex canny DeepLabv3+ FusionNet for enhanced detection and classification of bleeding regions in GI tract

## TL;DR

A new AI framework improves detection and classification of bleeding in gastrointestinal images, helping doctors diagnose and treat more effectively.

## Contribution

A novel multi-stage framework combining preprocessing, segmentation, and classification for enhanced bleeding detection in WCE images.

## Key findings

- The proposed framework achieved 97.6% mean pixel accuracy in segmentation.
- Classification accuracy reached 99.2% using ResNet-NaiveBayes Fusion.
- Dice Similarity Coefficient of 99.6% demonstrates high segmentation precision.

## Abstract

Detection of gastrointestinal bleeding in Wireless Capsule Endoscopy (WCE) images and accurate bleeding region segmentation and classification is crucial for exact diagnosis and treatment, as early detection can prevent severe complications. However, it remains challenging due to the inability of current methods to effectively differentiate between types of bleeding and handle complex borders of lesions. In this paper, a new framework: Multi-Stage Convo-Enhanced Retinex Canny DeepLabV3+ FusionNet is proposed to better tackle these challenges. Existing feature extraction algorithms struggle with colour differentiation and texture recognition, often failing to miss fine-scale textures that distinguish active bleeding from coagulated blood effectively. Hence, this approach is initiated with Clip-BiRetinexNet for preprocessing, enhancing image contrast and color consistency using Clipped Histogram Equalization and Bilateral Filtered Retinex thereby capturing fine-scale textures that distinguish active bleeding from coagulated blood. Existing segmentation and classification methods struggle with irregular and complex borders of bleeding types like ulcers and vascular lesions due to ineffective border detection. Therefore, the segmentation in this proposed model is handled by Hough Canny-Frangi Enhanced DeepLabV3+, improving edge detection and vascular pattern enhancement to delineate accurately irregular lesions. Next, a ResNet-NaiveBayes Fusion was shown for classification, offering effective probabilistic classification. The implementation results show that the proposed approach outperforms the state-of-the-art methods with a high mean pixel accuracy of 97.6%, classification accuracy of 99.2% and Dice Similarity Coefficient of 99.6%.