what is on-site retuning?

It means retraining a model where it is used, on images collected there, instead of shipping the data to a central lab and waiting. In this study the team labelled day one images, retrained for about two hours that evening, and put the new model back on the microscope for day two.

can I copy the model from the paper?

No. The paper says the data and algorithms may be made available on request, and gives no public repository. You do not need it. The service is tuning the model your client already owns, on their images.

why charge the device maker and not the lab?

The lab has no budget and did not choose the device. The group that paid for the tool is the one losing money while it sits unused. Sell to the health programme, the ministry, or the company that makes the microscope.

how good does the tuned model get?

In this study, on the same day two samples, brightfield catch rate reached 98.1 percent and clean rate reached 100 percent. Against ordinary light microscopy the tuned model caught 97.8 percent of infections but called only 68.2 percent of clean samples clean. Quote the second pair of numbers, not the first.

Sell a two-day retune to labs whose imported AI does not work

The finding this rests on

A team working near Azaguie, in rural Cote d’Ivoire, ran a urine testing clinic for two days. Their job was to find the eggs of a water-borne worm under a portable microscope. On day one the model that read the images caught 75.5 percent of the infected samples. Worse, it called only 46.7 percent of the clean samples clean.

Between the two days they fixed it on site. Seven people spent four hours marking up images from 36 samples, making 7,900 marks. Retraining took about two hours. The new model went back on the microscope the next morning.

On day two, on the same samples, the catch rate was 98.1 percent and the clean rate was 100 percent. Read the full write-up at Retuning AI on site raised parasite detection to 98.1%.

We score that paper 4 out of 10. The teams are strong and the field work is real. But it is a preprint, it covers about 100 people at one site, and the code and data are not public. So this idea is a bet that four hours of local labelling does the same thing somewhere else.

Who has this problem

Picture a lab manager at a district hospital in a rural part of Cote d’Ivoire. A donor programme gave her a portable microscope with a model inside it. The model was trained somewhere else, on samples that were collected somewhere else.

Her day runs like this. Forty people come in. The machine reads each slide and flags most of them as positive. She knows the flags are wrong, because she has looked down a microscope for eleven years. So she rechecks every slide by hand.

The machine has now made her slower, not faster. She stops using it. Six months later a visitor from the donor programme asks why the device is in a cupboard.

The person who feels this most is not her. It is whoever bought the device. That is a health programme, a ministry, or the company that makes the microscope. They paid for a tool that is sitting in a cupboard.

What you would sell them

Sell a two-day visit that retunes their model on their own samples. You bring one laptop, a labelling tool, and a person who knows how to train a model. Day one you collect and mark the site’s own images with their staff, and overnight you retrain.

On the second morning the updated model is back on their machine, reading real slides. You leave behind a one-page report with four numbers. The catch rate and the clean rate before, and the catch rate and the clean rate after.

What to charge

Test it at 4,000 dollars for one site visit. This is a hypothesis to check, not a price you can defend yet.

Here is the reasoning. Do not compare it to software. Compare it to a trip. Ask the buyer what it cost to fly one engineer to that site the last time the device broke. Ask what the device itself cost. Your fee should sit below the trip, and far below the device, because you are rescuing something they already own.

Test a second shape too. Offer three visits over a year for a lower price per visit, because a model that drifts once will drift again. Both prices are guesses. Ask for a purchase order, not for interest.

How you would build the first version

You are not building a product. You are building a repeatable two days.

Use an open source image labelling tool as it comes. Use one laptop that can train a small model in an evening. Write the retraining script once and reuse it. Write the report as one page in a spreadsheet, with the two error rates before and after, and nothing else.

Do not build a microscope. Do not build your own model. Do not build a cloud service, a dashboard, or a phone app. Note that this paper’s code and data are not public, so you cannot copy their model anyway. You do not need to. The whole job is tuning the model your client already has.

One thing you must sort out first. Get written permission to touch the client’s model, in an email, before you book a flight.

The one-week test

Day one. List every group in one country that has put an image reading diagnostic device into a rural lab. Aim for ten names.
Day two. Call five of them. Ask one question. How much did the clean rate drop when you moved the device to the field?
Day three. Rehearse the whole two days on public medical images. Time the labelling, time the retraining, and print the one-page report.
Day four. Send that report to the five groups you called. Offer the first visit at 4,000 dollars, with a date.
Day five. Ask each of them for a signed order or a deposit.

The answer is yes or no by Friday evening. Yes means one group has signed or paid. No means everybody wants to see more data first. Wanting more data first is a no.

What would kill this

The result does not travel. The model was tuned on images from one community and then tested on that same community the next day. About 100 people took part, at one site, in one month. Nothing in the paper shows what happens in the next district. If the gain shrinks by half elsewhere, your report will not sell a second visit.
Tuned may still not be good enough. Measured against ordinary light microscopy, the tuned model caught 97.8 percent of infections but called only 68.2 percent of clean samples clean. About three clean samples in ten were still flagged. The authors say this is close to meeting World Health Organization requirements, which means it does not meet them yet. Your client may pay you and still not be able to use the device.
You need an expert to label. The four hour session used seven people and produced 7,900 marks. Somebody in that room had to know what an egg looks like. At a site with one overworked technician, that labour does not exist, and you cannot supply it from another country.
Nobody may be allowed to change the model. Find out who signs off on a change to a diagnostic tool after it is installed. If the answer is a regulator, your two-day visit turns into a year of paperwork and your price is wrong.
It sells once. You fix the model, you leave, and the problem is gone. A business made of one-off visits is a job, not a company, unless drift brings you back.

Just Out Tech explains new research in plain language. This article was drafted with AI assistance and checked by a human against the original source.