Nowroozzadeh, M. H. | bioRxiv Science

bioRxiv2025

Attcatvgg-Net: an Explainable Multioutput Deep Learning Framework for Cataract Stage Classification and Visual Acuity Regression using Multicolor Fundus Images

PurposeThe purpose of this study is to develop and evaluate an attention-guided deep learning model using the multicolor imaging module of Spectralis Optical Coherence Tomography (OCT) imaging for automated cataract severity classification and Visual Acuity (VA) prediction. MethodsWe analyzed 314 multicolor fundus images from 169 patients. Images were preprocessed using an enhanced Retinex algorithm and segmented into three concentric macular zones: Zone 1 (fovea, central 1.5 mm diameter), Zone 2 (parafovea, 1.5-2.5 mm ring), and Zone 3 (perifovea, >2.5 mm radius). A multi-output convolutional neural network (AttCatVgg-Net), based on VGG-16 and enhanced with a Convolutional Block Attention Module (CBAM), was trained to simultaneously perform three-class cataract classification (normal to mild, moderate, severe) and visual acuity (VA) regression. Model performance was assessed using accuracy, AUC, F1-score, and regression metrics. Statistical analyses included the Wilcoxon signed-rank test and the Spearman correlation test. ResultsFor cataract grading, the integrated model using all wavelengths and zones achieved 92.5% accuracy, 94.7% area under the ROC curve (AUC), and a 92.1% F1-score. The green channel alone achieved 90.1% accuracy and 0.93 AUC, while the red channel yielded lower performance (76.3% accuracy, 0.83 AUC). Among anatomical zones, Zone 1 (fovea) and Zone 3 achieved 84.3% and 84.71% accuracy and 0.88 and 0.89 AUC, respectively, whereas Zone 2 underperformed (60.41% accuracy, 0.71 AUC). For visual acuity prediction, the full model achieved a mean absolute error (MAE) of 0.1181 and a coefficient of determination (R-squared) of 0.7759. The green channel demonstrated the strongest correlation with actual VA (correlation coefficient = 0.823, p < 0.001), followed by green-red (0.817) and blue (0.809). The green channel also achieved the lowest Mean Squared Error (MSE = 0.0369) and Root Mean Square Error (RMSE = 0.1920), outperforming other channels. ConclusionsAttention-guided deep learning applied to Spectralis OCT multicolor imaging enables accurate, objective classification of cataract severity and estimation of cataract-related visual acuity loss.

bioengineering↗

bioRxiv2025

RPEGENE-Net: A Multi-Resolution Deep Learning Framework for Predicting Gene Expression from Microscopy Images of Retinal Pigment Epithelium (RPE) Cells

PurposeTo develop a deep learning framework, RPEGENE-Net, capable of predicting gene expression profiles of retinal pigment epithelium (RPE) cells using live-cell microscopy images. MethodsA dataset of live-cell images of RPE cells, treated with various drug regimens and captured at magnifications of 40x, 100x, 200x, and 400x, was used. Gene expression of six key genes involved in epithelial-mesenchymal transition or EMT (including -SMA, ZEB1, TGF-{beta}, CD90, {beta}-catenin, Snail) and treatment classes (Aflibercept, Bevacizumab, Dexamethasone, Aflibercept + Dexamethasone, and untreated control) were analyzed. After preprocessing the image data and gene expression values, we trained and evaluated twelve state-of-the-art deep learning architectures, including three variants of DenseNet, five variants of ResNet, EfficientNet_b5, Inception_v3, RegNet_y_400mf, and a vision transformer model (Swin_b). A two-stage pipeline was implemented, combining autoencoder-based pretraining to extract meaningful features with fine-tuning specifically optimized for gene expression regression and treatment type classification tasks. Features extracted from the second stage across four magnifications were concatenated to generate the final prediction, leveraging multi-scale morphological information for improved accuracy. ResultsDenseNet121 demonstrated superior performance, achieving the highest Pearson correlation coefficients for four genes: -SMA (0.79), ZEB1(0.84), TGF-{beta} (0.83), and Snail (0.86). ResNet34 outperformed other models for CD90 (0.87) and {beta}-catenin (0.85) predictions. The average mean absolute error (MAE) and average root mean square error (RMSE) on test dataset were 0.0244 and 0.1228, respectively. The R2 scores ranged from 0.50 (-SMA) to 0.74 (TGF-{beta}), indicating strong alignment between predicted and actual gene expression values. A multi-level approach, combining data from 40x, 100x, 200x, and 400x magnifications yielded higher R2 scores for almost all genes compared to single-magnification models. For the classification task, DenseNet121 achieved F1 score, precision, recall, and accuracy of 0.98, with a specificity of 0.99. ConclusionsRPEGENE-Net provides a simple, cost-effective method to predict gene expression from live-cell images, with potential applications in experimental studies, RPE transplantation quality control, and broader cell-based research. Multi-magnification imaging enhances model performance, supporting its utility as a scalable tool for diverse gene expression studies.

bioengineering↗