Subject lists
Load normal and glaucoma subject lists for macular and optic-disc scans.
Project · Medical imaging
OCT Processing Pipeline
A placeholder project page for preparing OCT-derived RNFL and GCIPL maps for glaucoma-focused representation learning.
Summary
Why preprocessing matters.
OCT-derived retinal maps are not immediately model-ready. They require consistent file discovery, class membership handling, visit pairing, anatomical alignment, normalisation, masking, and quality checks.
This project page is a placeholder for documenting the practical workflow that transforms raw or semi-processed OCT-derived maps into clean input tensors for MixtureBetaVAE-style modelling.
Later, this page can include real examples of RNFL and GCIPL maps, pairing statistics, cohort summaries, preprocessing scripts, and reconstruction examples.
Workflow modules
From file discovery to model input.
File parsing
Recover patient ID, scan type, date, eye, serial number, and map type from filenames.
Visit pairing
Pair ONH RNFL and macular GCIPL maps using exact or fallback visit-date matching.
Anatomical alignment
Flip right-eye maps horizontally so anatomy is consistently aligned across eyes.
Normalisation
Scale thickness values into a stable model range while preserving clinical meaning.
Masking and losses
Use masks or region weights to handle optic-disc areas and clinically important structures.
Conceptual workflow
From OCT files to paired maps.
OCT Files
→
Parse Metadata
→
Pair Visits
→
Align + Normalise
→
Model Input
Placeholder note:
Replace this section with real diagrams, manifest statistics, and representative paired examples later.
Placeholder outputs
Artifacts to add later.
RNFL map
Placeholder for optic-disc RNFL thickness map examples.
GCIPL map
Placeholder for macular GCIPL thickness map examples.
Pairing manifest
Placeholder for CSV summaries, class balance, and paired visit counts.
Technical reflection
Dataset quality shapes model quality.
Class balance matters.
Paired multi-modal data can introduce severe imbalance, especially when exact visit-date matching is required.
Alignment is part of the model design.
Anatomical orientation decisions affect whether the model learns meaningful structure or avoidable left-right variation.
Loss weighting should reflect the domain.
Reconstruction objectives can be improved by weighting clinically important regions rather than treating all pixels equally.