Retinal Imaging for Eye Disease

Fundus photos and OCT slices show vessels and layers where tiny bleeds and fluid mark diabetic disease early.

Why Does This Exist?

Diabetes blinds through tiny vessel leaks years before vision blurs. Fundus cameras shoot the retina through the pupil in seconds; OCT sweeps a laser to slice layers in microns. Input is a dilated photo plus optional OCT cube; output is referral for lesion detection and fluid. Cataracts cloud the view and break the chain.

Think of It Like This

Aerial photos plus core samples of a marsh

Aerial photos spot dark pools while core samples show water depth by soil layer. Pools alone cannot tell depth, and cores alone miss the map. Together they decide which fields drain this week.

Fundus is the aerial and OCT is the core. The analogy stops at fixation: eyes saccade while marshes sit still.

How It Actually Works

Fundus frames vessels near $100$ microns wide; microaneurysms dot $20$ to $100$ microns. OCT resolves $5$ -micron layers; fluid thickens the macula past $300$ microns from a normal $250$ . Grading counts bleeds per quadrant plus fluid location relative to the fovea.

Worked example

A photo shows $12$ microaneurysms and $3$ blot bleeds in two quadrants, which triages moderate disease. OCT measures central thickness $340$ microns, $90$ over normal, with fluid $200$ microns from the fovea. That distance plus thickness refers within days, while bleeds alone would watch monthly.

Watch Out For

Cataract haze hiding bleeds

A cloudy lens drops contrast $30$ percent so $50$ -micron dots vanish. The symptom is clean AI reads on unreadable photos. Grade image quality first and dilate or refer when vessels blur.

The Quick Version

Fundus photos map bleeds fast; OCT slices layers for fluid depth.
Hundred-micron dots and $300$ -micron thickness set referral clocks.
Location relative to the fovea matters more than raw counts.
Do not trust clean reads on hazy cataract photos.
Dilation plus fixation decide quality before any model runs.