RFMiD 2.0: Evaluation Metrics for a Lightweight CNN on Retinal Disease Images

Evaluation metrics for the LiteFeatNet convolutional neural network tested on the RFMiD 2.0 dataset. The model, developed by Usman Rafi, was trained on 1,824 retinal fundus images across three disease classes and achieved a top testing accuracy of 90.33%. The dataset was last updated on May 11, 2026.

Use Cases

• Benchmarking lightweight CNN architectures based on reported accuracy, precision, recall, and F1-score metrics.
• Assessing model inference speed for clinical deployment based on the reported 4-millisecond inference time.
• Evaluating model generalizability using external validation datasets like RFMiD 2.0.
• Conducting ablation studies on feature extraction and refinement modules based on the described methodology.

Strengths

• Model performance is quantified with specific metrics: 90.33% accuracy, 90.69% precision, 90.33% recall, and 90.27% F1-score.
• The model's compact size of 19.87 MB and fast inference time of 4 milliseconds per image are documented.
• Experiments were conducted using 1,824 images from the RFMiD dataset with a defined 60:20:20 train-validation-test split.

Limitations

• Row count is unknown, which may limit suitability assessment.
• Column-level documentation is absent; field semantics must be inferred after download.
• The dataset is very small at 9.5 KB, suggesting it contains only summary evaluation metrics, not the underlying image data.

Provenance

Source

figshare

Collection Method

Likely contains performance metrics from model training and evaluation experiments.

Freshness

Last updated 2026-05-11 17:32:51; freshness should be verified.