Cataract Diagnostics
Smartphone-based visual health screening. Follow the alignment reticle overlays using your device camera or upload a raw eye image file below to run analysis.
Upload Eye Scan
Drag & drop your ophthalmic scan image here, or select a file to upload.
Awaiting Diagnostics Scan
Acquire Macro Pupil Focus
Align the patient's iris under uniform, diffuse ambient lighting. Avoid bright specular glare on the cornea.
Feature Map Extraction
A compressed 8-layer Vision Transformer (DINOv2) processes the optical image to extract key structural parameters.
Support Vector Classification
Features are compressed via PCA to 64 components and classified by an RBF SVM to evaluate lens density.
Compressed QAT DINOv2 Pipeline
Achieving over 96% binary classification accuracy by compressing transformer weights down to 30.4MB for lightweight edge and CPU environments.
Structured Pruning
Dropped the final 4 layers of the standard 12-block DINOv2-small transformer, decreasing parameters to 15.0M while retaining 99% of crucial early edge and pattern sensitivities.
Quantization-Aware Training
Transformer MLP linear projections were fine-tuned with fake-quantization. Post-training static conversion maps weights to INT8, shrinking memory footprint to 30.4MB.
PCA + SVM Inference
High-dimensional 384D embedding vectors are projected via PCA to 64 dimensions to reduce noise. An RBF SVM classifier computes the final diagnostic predictions.