Region Extraction with Deep Cascaded Neural Architecture for Preserving Image Intensities
Lists
Only you see this
Abstract
Deep learning (DL) has recently achieved a massive response for medical image analysis, especially in semantic image segmentation. Semantic neural architectures are pivotal in extracting anatomical structures from provided input data. The conventional models, inspired by U-Net architecture with encoder-decoder structure, encounter irreversible intensity loss. It is due to continuous stride and pooling operations that prevent them from precisely extracting vascular tree structure and impact the efficacy of degenerative retinal disease detection. In this work, we proposed a Cascaded Neural Architecture (CNA) that uses a generative adversarial concept with two generators and discriminators to localize coarse and fine features. The network helps in preserving vessels by classifying input at pixel-level. In generators, we introduced residual blocks with a densely connected residual block used only in the bottleneck part of a fine-generator for maximum feature reuse and to restrain redundant feature sharing. We also designed customized lightweight discriminators with encoder-decoder structures to reduce computable parameters. Our network is compared with prominent architectures and evaluated on publicly available fundus image datasets (DRIVE, STARE, and CHASE-DB1) with variations in dimensions and focal points. Our network has achieved promising results by reporting the highest area under the ROC curve and structural similarity index score.
BibTeX
@inproceedings{khan2022region,
author = {Muhammad Zubair Khan and Yugyung Lee},
title = {{Region Extraction with Deep Cascaded Neural Architecture for Preserving Image Intensities}},
booktitle = {BIBM},
year = {2022},
url = {https://openreview.net/forum?id=lSjCglsmVpT},
}Loading related content...