Abstract

Info

Title:: Automated detection of glaucoma using recursive feature elimination with cross-validation and hybrid neural decision forest

Author(s):: SHANG Linan1; 2; CHI Yue1; 2; ZHOU Yatong1; 2; SHAN Chunyan3; XIAO Zhitao4; JIN Kezhen1; 2; 1. School of Electronic and Information Engineering, Hebei University of Technology, Tianjin 400401, China 2. Institute for Innovation, Hebei University of Technology, Shijiazhuang 050299, China 3. Tianjin Medical University Chu Hsien-I Memorial Hospital, Tianjin 300134, China 4. School of Life Sciences, Tiangong University, Tianjin 300387, China

Keywords:: Keywords: glaucoma Vision Transformer hybrid neural decision forest recursive feature elimination with cross-validation

Abstract:: Abstract: This study proposes an automated glaucoma detection model called VITRCH which integrates recursive feature elimination with cross-validation (RFECV) and a hybrid neural decision forest (HNDF). To better capture global visual information from retinal fundus images, a Vision Transformer (ViT) model is employed for retinal fundus image classification. Then, the multi-layer perceptron classifier in the feature classification branch of the original ViT model is replaced by an HNDF classifier which takes the advantage of the stronger feature representation ability of neural networks together with the interpretability and robustness of decision forests, thereby enabling more efficient glaucoma image classification. Subsequently, an RFECV module is incorporated after the feature extraction stage of the original ViT to reduce training complexity and shorten training time by eliminating redundant features with low contribution rates to the classification tasks. The improved models are evaluated on the ACRIMA dataset, and experimental results demonstrate that, compared with the original ViT model, the VITRCH model effectively leverages the advantages of each improved module and provides a more robust solution for automated glaucoma detection.