What is it about?

Diagnosing Gastrointestinal Diseases from Endoscopy Images through a Multi-Fused CNN with Auxiliary Layers, Alpha Dropouts, and a Fusion Residual Block. In this work, a proposed Multi-Fused Residual Convolutional Neural Network (MFuRe-CNN) with Auxiliary Fusing Layers (AuxFL), a Fusion Residual Block (FuRB) both with Alpha Dropouts (αDO) diagnosed various endoscopic images of gastrointestinal (GI) conditions or diseases. The proposed MFuRe-CNN handled four cases, including colons with ulcerative colitis, polyps, esophagitis, and a healthy colon, sourced from reliable databases like the KVASIR and ETIS-Larib Polyp DB. The proposed model consisted of three state-of-the-art models fused into a single feature extraction pipeline with their partially frozen and truncated layers, which helped propagate robust features and improved the diagnostic performance without consuming a hefty fraction of computing cost compared to most existing state-of-the-art models. In addition, the MFuRe-CNN incorporated with AuxFLs, αDOs, and FuRB have shown a significant contribution in reducing overfitting and performance saturation compared to those without the said components. Upon evaluation, the proposed model achieved an outstanding 97.25% test accuracy with only 4.8 M parameters and consumed 7.8 GFLOPs during inference, making it more efficient and accurate than most conventionally trained DCNNs. As concluded, the proposed MFuRe-CNN can potentially improve the diagnosis of the GI tract more cost-efficiently than ensembles and perform better diagnosis than most conventional pre-trained and fine-tuned DCNNs.

Featured Image

Why is it important?

Highlights •The MFuRe-CNN performed better than most DCNNs. •Using AuxFLs, αDropouts, and FuRB improved the overall performance. •Cost-efficient, straightforward, and trains only on a single pipeline. •A fused model that requires only 4.8M parameters to train.

Read the Original

This page is a summary of: Diagnosing gastrointestinal diseases from endoscopy images through a multi-fused CNN with auxiliary layers, alpha dropouts, and a fusion residual block, Biomedical Signal Processing and Control, July 2022, Elsevier, DOI: 10.1016/j.bspc.2022.103683.
You can read the full text:




The following have contributed to this page