Enhanced Detection of Colon Diseases via a Fused Deep Learning Model with an Auxiliary Fusion Layer and Residual Blocks on Endoscopic Images.

Abstract

Background: Colon diseases are major global health issues that often require early detection and correct diagnosis to be effectively treated. Deep learning approaches and recent developments in medical imaging have demonstrated promise in increasing diagnostic accuracy.

Objective: This work suggests that a Convolutional Neural Network (CNN) model paired with other models can detect different gastrointestinal (GI) abnormalities or diseases from endoscopic images via the fusion of residual blocks, including alpha dropouts (αDO) and auxiliary fusing layers.

Methods: To automatically diagnose colon disorders from medical images, this work explores the use of a fused deeplearning model that incorporates the EfficientNetB0, MobileNetV2, and ResNet50V2 architectures. By integrating these features, the fused model aims to improve the classification accuracy and robustness for various colon diseases. The proposed model incorporates an auxiliary fusion layer and a fusion residual block. By combining diverse features through an auxiliary fusion layer, the network can create more comprehensive and richer representations, capturing intricate patterns that might be missed by single-source processing. The fusion residual block incorporates residual connections, which help mitigate the vanishing gradient problem. By adding the input of the block directly to its output, these connections facilitate better gradient flow during backpropagation, allowing for deeper and more stable training. A wide range of endoscopic images are used to assess the proposed model, offering an accurate depiction of various disease scenarios. Results The proposed model, with an auxiliary fusion layer and residual blocks, exhibited an enormous reduction in overfitting and performance saturation. The proposed model achieved an impressive 98.03% training accuracy and 97.90% validation accuracy after evaluation, outperforming the majority of typically trained DCNNs in terms of efficiency and accuracy.

Conclusion: The proposed method developed a lightweight model that correctly identifies disorders of the gastrointestinal (GI) tract by combining advanced techniques, including feature fusion, residual learning, and self-normalization.