Internship Proposal: MLP-Based Pathology Classification and Stratification from Retinal Arterial and Related Ocular Pulse Waveforms
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
Beat-resolved retinal arterial velocity waveforms measured by laser Doppler holography (LDH)carry information about pulsatility, damping, vascular loading, and the cumulative functionalstate of the downstream microvascular bed. Recent work has formalized a clinician-orientedlibrary of explicit, auditable, and transportable waveform-shape metrics designed to remaininterpretable even when absolute velocity calibration is uncertain. Building on this foundation,This internship proposes a waveform-level machine-learning pipeline in which a multilayerperceptron (MLP) takes a normalized pulse waveform as input and predicts eitherpathology class or a continuous disease-severity score. The primary focus is on retinalarterial waveforms, but we will also test whether choroidal arterial waveforms and retinaland choroidal venous waveforms provide complementary predictive information. The core idea is to work directly from the pulse shape rather than only from pre-aggregatedbiomarker vectors. Each beat-resolved waveform is resampled onto a common cardiac phase grid,normalized to reduce sensitivity to unknown amplitude gain, and passed to an MLP that learnsnonlinear combinations of waveform morphology associated with healthy and pathologicalmicrovascular states. Binary diagnosis, multiclass staging, and continuous stratification will beevaluated at the beat, vessel, and eye levels. The same framework can optionally ingest sideinformation such as age, blood pressure, intraocular pressure, and selected interpretablewaveform-shape metrics. A central objective is to preserve clinical transparency despite using a neural model. We willtherefore combine MLP prediction with interpretable summaries derived from the waveform-metriclibrary, phase-wise saliency analyses, confidence estimates, and robust aggregation across beats.The expected outcome is a reproducible and physiologically grounded framework for waveform-basedretinal pathology classification that is suitable for longitudinal and multi-centre studies,with an initial emphasis on retinal arterial inputs and optional extension to choroidal arterialand retinal or choroidal venous inputs.
Fields of Study & Subjects
Retinal Imaging and Analysis
Artificial Intelligence in Healthcare
Non-Invasive Vital Sign Monitoring
Publication details
- Journal / source
- Zenodo (CERN European Organization for Nuclear Research)
- Publication date
- 30 March 2026
- Language
- EN