Retinal image processing pipeline and an encrypted on-device biometric engine
Biomedical image and signal processing · Reproducible computer vision and on-device biometric security
A processing pipeline that isolates the blood vessels in retinal fundus images, paired with an encrypted on-device engine for biometric records and schedules. This system is not a medical device and is not intended for clinical diagnosis.
The challenge
Retinal fundus imagery suffers from signal-dependent Poisson-Gaussian CMOS sensor noise and uneven spherical illumination. Isolating micro-vascular trees and extracting feature vectors requires a strictly deterministic mathematical pipeline that produces bit-identical masks across hardware architectures without cloud wrappers.
Paired with an on-device biometric and schedule engine, all sensitive records and processing outputs must remain encrypted at rest on the local device using page-level authenticated encryption with zero external network transmission.
- Sensor noise
- Generalised Anscombe variance stabilisation with exact unbiased inverse
- Vessel extraction
- Multi-angle morphological top-hat and analytic 2×2 Hessian Frangi filters
- Reproducibility
- Bit-identical IEEE 754 floating-point masks across architectures
- On-device security
- Page-level AES-256-GCM authenticated storage with zero network I/O
Architecture
First-principles numerical pipeline
Spatial filtering, noise variance stabilisation, morphological transformations, and matrix decompositions are implemented directly over contiguous multi-dimensional arrays, operating completely offline without cloud APIs or opaque third-party wrappers.
Bit-identical cross-platform reproducibility
The floating-point pipeline is strictly deterministic under IEEE 754 rules, producing bit-for-bit identical vascular segmentation masks and feature vectors across different processor architectures.
Raw CMOS retinal buffers pass through generalised Anscombe noise stabilisation and green-channel CLAHE into multi-angle morphological top-hat filtering and 2×2 Hessian eigenvalue vessel segmentation, producing bit-identical masks stored locally inside an AES-256-GCM page-encrypted vault with zero network traffic.
Hardware and protocols
Sensor noise stabilisation and adaptive contrast
A generalised Anscombe transformation converts the CMOS sensor's signal-dependent Poisson-Gaussian noise into uniform Gaussian noise, paired with an exact unbiased inverse transform. The green channel is then extracted and enhanced via contrast-limited adaptive histogram equalisation (CLAHE) with bilinear tile interpolation.
Multi-scale morphological and Hessian vessel segmentation
The retinal vascular tree is isolated by combining multi-angle linear morphological top-hat transforms with multi-scale Frangi vesselness filtering derived from the analytic eigenvalues of the local 2×2 Hessian curvature matrix.
Page-encrypted on-device biometric engine
Biometric records and medication schedules are managed by a local-only engine that encrypts each storage page using authenticated AES-256-GCM encryption, executes deterministic local scheduling state machines, and transmits zero data over external networks.
Verification
A self-contained numerical suite generates synthetic retinal phantoms at test time: Bézier vessel trees, Gaussian blobs and calibrated Poisson-Gaussian sensor noise.
Automated checks confirm the signal-to-noise gain, a Dice similarity above 0.92 against the synthetic ground-truth masks, and rotational invariance, and need no external image dataset.