How we do it

The Science

Non-invasive digital ocular biomarker retinopathy, disease detection with HIPAA and GDPR compliance.

1
2
3
4
5
6
Download Clinical Trial Doc
t
1

Custom biomedical imaging modules to capture high-resolution retinal images. Precision-matched LEDs (530–600 nm) and aspheric lenses optimize light capture and reduce distortion. Lens-sensor design guided by optical simulation to achieve <10 µm resolution. A microcontroller-driven illumination ring enables polarization and multispectral imaging for enhanced contrast in vascular and melanin-rich tissues.

t
2

Cipher applies an advanced domain-specific preprocessing before AI inference. For the captured retinal images, it uses vessel enhancement (e.g., Frangi filters), green-channel extraction, and CLAHE to highlight features like microaneurysms. We use shape analysis to support classical and deep learning models. Noise and lighting variations are corrected using flat-field and Retinex-based adjustments..

t
3

Cipher’s AI backbone includes deep convolutional neural networks tailored to each diagnostic task:

  • Fundus classification
  • Cellular object detection
    • Using heavy data augmentation (elastic deformation, Gaussian blur, color jitter) The networks are trained using focal loss and label smoothing to address class imbalance, with augmentation guided by CutMix and RandAugment techniques. Performance is validated with stratified k-fold cross-validation and ROC-AUC optimization.

    t
    4

    To meet power and compute constraints, models are compressed using:

  • Quantization-aware training (QAT) for INT8 deployment
  • Pruning (structured and unstructured) to reduce parameter load
  • Knowledge distillation from full-sized teacher models into lightweight student networks
    • Optimizations ensure <200ms inference time, <200MB memory usage, and battery-viable power consumption profiles (<5W peak)

    t
    5

    Cipher integrates explainability methods to analyze model behavior post-hoc and support regulatory interpretability. Techniques used include:

  • Grad-CAM++ to localize pathology-relevant regions in fundus images
  • Integrated Gradients and LIME for model behavior approximation and sensitivity attribution
  • Class-discriminative localization maps (CDL) to visualize layer activation correlation with histopathological features
    • Feature importance heatmaps are benchmarked against human annotation masks using intersection-over-union (IoU) and pixel-level F1 scores to validate interpretability fidelity.

    t
    6

    Clinical validation of our AI predictions are anchored in multi-label agreement between model outputs and physician consensus labels. Ground truth generation involved triple-blinded expert annotation with inter-rater agreement (Cohen’s κ > 0.82). Signal alignment between captured images and biological markers was verified via:

  • Spectrophotometric analysis of image vs. hemoglobin content in blood smears
  • Vessel width-to-optic-disc ratios to confirm fundus image fidelity
  • Colorimetric lesion segmentation accuracy to confirm parasitaemia
    • All data pipelines are documented and aligned with Good Machine Learning Practice (GMLP) and ISO/IEC 23053:2022 for machine learning model lifecycle management in medical applications.