Automated cataract detection has long been a major focus of ophthalmic artificial intelligence research, but balancing diagnostic accuracy with computational efficiency remains a challenge. A new study published in Scientific Reports describes a hybrid deep learning framework that aims to address both – combining optimized feature selection with a lightweight yet high-performing classification network.
The authors propose a dual-component system: Chaotic Adaptive Poplar-Bacteria Optimization (Cha-PO) for feature selection, and Cataract VisionNet (CVNet) for classification. Together, the approach achieved 99.10 percent accuracy on the Eye Cataract Kaggle dataset, alongside 99 percent precision, 99.21 percent recall, and a 99.10 percent F1-score.
Fundus image-based cataract detection often relies on convolutional neural networks (CNNs). While powerful, there can be issues associated with this approach – standard CNN pipelines can suffer from feature redundancy, predisposition to overfitting, and high computational cost and complexity, particularly problematic in resource-limited or real-time screening environments.
The Cha-PO algorithm addresses this by performing optimized feature selection before classification. By integrating chaotic mapping with adaptive bacterial foraging and poplar tree-inspired search mechanisms, Cha-PO aims to reduce dimensionality while preserving diagnostically relevant features. The result is fewer redundant parameters and improved computational efficiency. From a clinical standpoint, this is particularly relevant in teleophthalmology and mobile screening contexts, where hardware constraints and scalability are critical considerations.
The second component of the hybrid approach, CVNet, is a transfer learning–based architecture that combines GhostNet, Inception-v3, and NASNet models. The design incorporates lightweight convolutional layers, multiscale feature extraction, and attention-based refinement to enhance discriminative performance while maintaining efficiency.
The model was trained and validated on fundus photographs representing varying cataract severities – from normal to mild, moderate, and severe cases. Preprocessing steps included normalization, contrast enhancement, and grayscale segmentation to improve visualization of cataract-related features.
Compared with traditional machine learning methods – K-Nearest Neighbors (KNN), Logistic Regression (LR), Decision Tree (DT) – the proposed approach significantly outperformed all benchmarks. Accuracy improved to 99.10 percent, versus 80–87 percent for conventional classifiers.
The model also surpassed widely used deep learning architectures including VGG19, ResNet50, DenseNet201, InceptionV3, and EfficientNetB0, achieving approximately 3–5 percent higher accuracy while maintaining substantially lower computational cost.
The study authors suggest that the system could be extended beyond cataract detection to grading severity, monitoring progression, and potentially predicting cataract surgery outcomes.
However, as with all AI-based diagnostic tools, external validation across diverse populations and imaging systems remains essential before clinical deployment. If validated in broader datasets, such hybrid approaches may help scale cataract screening while maintaining the diagnostic standards clinicians have come to expect.
The authors propose a dual-component system: Chaotic Adaptive Poplar-Bacteria Optimization (Cha-PO) for feature selection, and Cataract VisionNet (CVNet) for classification. Together, the approach achieved 99.10 percent accuracy on the Eye Cataract Kaggle dataset, alongside 99 percent precision, 99.21 percent recall, and a 99.10 percent F1-score.
Fundus image-based cataract detection often relies on convolutional neural networks (CNNs). While powerful, there can be issues associated with this approach – standard CNN pipelines can suffer from feature redundancy, predisposition to overfitting, and high computational cost and complexity, particularly problematic in resource-limited or real-time screening environments.
The Cha-PO algorithm addresses this by performing optimized feature selection before classification. By integrating chaotic mapping with adaptive bacterial foraging and poplar tree-inspired search mechanisms, Cha-PO aims to reduce dimensionality while preserving diagnostically relevant features. The result is fewer redundant parameters and improved computational efficiency. From a clinical standpoint, this is particularly relevant in teleophthalmology and mobile screening contexts, where hardware constraints and scalability are critical considerations.
The second component of the hybrid approach, CVNet, is a transfer learning–based architecture that combines GhostNet, Inception-v3, and NASNet models. The design incorporates lightweight convolutional layers, multiscale feature extraction, and attention-based refinement to enhance discriminative performance while maintaining efficiency.
The model was trained and validated on fundus photographs representing varying cataract severities – from normal to mild, moderate, and severe cases. Preprocessing steps included normalization, contrast enhancement, and grayscale segmentation to improve visualization of cataract-related features.
Compared with traditional machine learning methods – K-Nearest Neighbors (KNN), Logistic Regression (LR), Decision Tree (DT) – the proposed approach significantly outperformed all benchmarks. Accuracy improved to 99.10 percent, versus 80–87 percent for conventional classifiers.
The model also surpassed widely used deep learning architectures including VGG19, ResNet50, DenseNet201, InceptionV3, and EfficientNetB0, achieving approximately 3–5 percent higher accuracy while maintaining substantially lower computational cost.
The study authors suggest that the system could be extended beyond cataract detection to grading severity, monitoring progression, and potentially predicting cataract surgery outcomes.
However, as with all AI-based diagnostic tools, external validation across diverse populations and imaging systems remains essential before clinical deployment. If validated in broader datasets, such hybrid approaches may help scale cataract screening while maintaining the diagnostic standards clinicians have come to expect.
