Cataract surgery is the most common operation on Earth — 27 million patients a year need it to keep their sight. And the math behind it is moving in the wrong direction: demand is climbing while the surgical workforce shrinks. Every future surgeon has to get good faster, with less senior time available to teach them.
+35%
demand in eye surgeries by 2030
from 700k/year today in Germany alone
−12%
ophthalmology workforce (US)
with comparable drivers in Germany
Sources: Berkowitz et al., Ophthalmology (2024), projection to 2035 · DOG (2025)
How does a resident get better today? An attending watches and gives notes. That system has three structural problems — and none of them get better by having fewer surgeons:
Senior-dependent
Review is entirely dependent on a senior physician being available — the scarcest resource in the building.
Manual, on top of surgery
Reviews happen manually, squeezed on top of full surgical schedules. Feedback arrives late or not at all.
Inconsistent metrics
Every reviewer measures differently — outcomes across trainees, hospitals, and time are simply not comparable.
With OcuFlow, a cataract surgery video is structured into its surgical phases automatically — incision, viscoelastic, capsulorhexis, phaco, and on — so feedback can be anchored to the exact moments that matter. And this part is not a mock: a frozen ImageNet ResNet50 feeding a custom MS-TCN head (the published TeCNO recipe), trained on Cataract-1K — no turnkey pretrained model exists for this task. Below: the unedited output for one held-out case. Drag through it.
model confidence 96.1%
Drag the strip. Gaps between segments are idle windows (instrument exchanges) the model correctly leaves unlabelled. This is unedited output of the 4-fold-CV model — the same segments the demo plays.
0.918 ± 0.025
validation accuracy, 4-fold CV over all 56 cases
0.871 ± 0.030
macro F1 — above the paper's 0.78–0.85 baselines
Deliberately the published recipe — no class weighting, no augmentation tricks — so the numbers are defensible against the literature, not inflated. Trained locally on a MacBook M2 Pro: ~85 minutes, $0 cloud compute.
Per-phase F1 highlights
long, visually distinctive
long, visually distinctive
long, visually distinctive
short, visually similar — the same hard case the original paper flags
Phases give feedback its where; instrument motion gives it its what. The pipeline computes, per phase, the motion metrics clinical studies already use for resident assessment — path length, velocity, direction changes, dwell time — plus SPARC, a single-number smoothness score from the movement-science literature.
It also computes per-frequency-band power (0.5–3 Hz, 4–7 Hz, 8–12 Hz). The hypothesis behind those features — that controlled fine-motor work and hand instability live in different bands — comes from the motion-analysis literature, and it stays a hypothesis until paired skill-score data validates it. Below is what can be shown honestly today: the real labelled tip track from case 5104, split with a simple low-pass into its smooth path and its high-frequency residual.
Pipeline report — Incision I
Every point is a real labelled position of the instrument tip (in-house web labeller + Lucas-Kanade propagation). "Smoothed" is a 0.4 s low-pass of the same track; "Micro-motion" colors that path by how much the low-pass removed at each moment — part hand motion, part tracking noise, not a validated tremor measurement.
Cataract-1K only ships sparse masks (~every 5th frame). Motion analysis needs every frame, so four tracking/labelling approaches were built and honestly compared:
AGT-seeded Lucas-Kanade pipeline
the workhorsePerceptual-hash GT frames back to their video position, seed tool tips, propagate LK re-anchoring drift at each GT frame. Batch-rendered tracked-motion video for 27 cases (3.3 GB).
BIn-house web keypoint labeller
built into the productManual sparse seeding + LK propagation + per-frame drag-to-correct, 5 API endpoints. Produced the 100%-coverage track powering the figure above.
CDUSTrack integration
research-grade fallbackWrapped a semi-automatic research tracker in an isolated conda env with a converter into the project’s track schema.
DCVAT
evaluated & rejectedDockerized, tested, documented why it isn’t enough: linear-only interpolation between keyframes can’t follow surgical motion.
Where this goes once paired skill-score data lands: every surgery a resident performs becomes a data point on their own curve, and every phase can be laid beside an expert's. Both views are built and interactive today — with illustrative data, shown exactly as they were pitched at the finale.
OcuFlow will track your progress…
Skill score and metric trends across your last 24 cases, per phase, against an expert benchmark — with a recommended focus phase where practice pays off most.
Open the progression view →…and enable you to learn from the best
Your surgery beside an expert benchmark, phase-locked step for step — both videos mapped to the same point in the phase, whatever pace each surgeon worked at.
Open the comparison view →Surgical-training software sits inside a $7.5B market. The wedge is a $9 pay-per-case entry — low enough for a resident to expense — growing into the real revenue driver: $15–30k/year SaaS per clinic. The positioning claim from the deck: existing options are either affordable but passive (content channels) or active but expensive (Zeiss, simulator hardware). OcuFlow takes the empty quadrant.
The finale closed on the ask: a €1.2M pre-seed and clinical pilot partners — pitched to 200+ people at the Ägyptisches Museum.
Sources: Grand View Research (2024/25); SOM = 2% of Europe SAM · CAGR 5.7% to 2033
OcuFlow was a three-person team inside the StartLabs × OneAim program — two months paced by six pitches, ending on the museum stage. The pitching, positioning, and research were shared. The shipped product was not: every line of the platform — frontend, backend, model training, labelling tools, deployment — is mine.
David Vogenauer
AI & Computer Vision
Research direction on the vision pipeline
Chantelle Davis
Business & Go-to-Market
Market sizing, positioning, clinical outreach
Andrei Zitti
Product & Engineering
Sole builder of everything that runs — including this page and the demo it embeds