How our scanner compares with other health apps
We ran 643 items through the scanner members actually use and got 90% of them right. Below, that sits beside what published research reports about AI food-photo apps — with every outside number linked to the paper it came from.
Read this first
Side by side
| What is being read | DoseRoutine, measured | Published research |
|---|---|---|
| Recognising a plate of mixed food | 88% of 493 label-checked real meal photos read correctly | Published app testing found automatic estimates from food photos unreliable on mixed and culturally diverse dishes, even for the two best-scoring apps (Nutrients 2024) |
| Recognising a single packaged product | 95% of 60 real packaged-food photos read correctly | Best research models passed 90% recall on single-product photos (CVPR 2025 workshop paper (arXiv preprint) 2025) |
| Reading a printed supplement, vitamin, peptide, TRT or prescription label | 99% of 80 labels with known contents read correctly | Not covered by the nutrition literature — diet apps do not read these labels (CVPR 2025 workshop paper (arXiv preprint) 2025) |
| Barcodes and QR codes | 10 of 10 decoded on the phone itself, with no AI cost | Standard across the category; barcode scanning is a solved problem (Nutrients 2024) |
| Telling you when it is unsure | Every read carries a confidence score, with thresholds set from measured accuracy | Reviewed studies report error only after the fact; confidence is rarely shown to the person using the app (Annals of Medicine (systematic review) 2023) |
Where the outside numbers come from
97% and 92% were the best food-image accuracies among seven AI nutrition apps
Researchers screened 800 nutrition apps in Australia and compared the output of seven AI food-image apps against weighed food records for Western, Asian and recommended diets. MyFitnessPal (97%) and Fastic (92%) came out highest, but the authors concluded that automatic energy estimates from food-image recognition were inaccurate overall, especially for mixed dishes and culturally diverse food.
Nutrients, 2024Over 90% recall for single-product photos from the strongest vision models
Six vision-language models were tested on 9,263 expert-labeled food images. The best passed 90% expert-weighted recall when a photo showed one clear product, and all of them dropped away on fine-grained distinctions such as cooking style and closely related foods.
CVPR 2025 workshop paper (arXiv preprint), 2025Across 52 studies, AI diet estimates missed the truth by anywhere from 0.1% to 38%
A systematic review of AI-based dietary assessment kept 52 peer-reviewed studies and found reported errors against ground truth spanning roughly 0.1% to 38%. The studies were too different from one another to pool, which is exactly why one app's headline percentage cannot be compared straight to another's.
Annals of Medicine (systematic review), 2023Independent accuracy testing of image-based diet apps is still a live research question
A 2024 evaluation in the Journal of the Academy of Nutrition and Dietetics assessed the accuracy of image-based dietary assessment mobile apps — further reading if you want the academic view rather than an app's own marketing.
Journal of the Academy of Nutrition and Dietetics, 2024Questions people ask
Is this a head-to-head test against MyFitnessPal or Cal AI?
No, and any app claiming that should be read carefully. Our numbers come from our own photo set; the other numbers come from published research using different photos and different ground truth. Comparing them tells you roughly where we sit in the category, not who wins a race.
How was the DoseRoutine number measured?
We ran 643 items through the same scanner members use: 493 real meal photos, 60 real packaged products, 80 printed supplement, vitamin, peptide, TRT and prescription labels with known contents, and 10 codes. Every meal photo's label was independently checked first, and 206 photos whose public caption did not match the plate were dropped rather than scored in our favour.
Why is meal recognition lower than label reading?
A printed label states exactly what is in the product. A plate does not: portion size, sauces and hidden ingredients have to be inferred from one photo. The research finds the same gap, which is why we show a confidence score on food photos instead of presenting every read as fact.
What happens when the scanner is unsure?
Below 70% confidence a food photo is sent for one harder second read before you see it. Anything still under 70% is flagged for you to check, and anything under 50% is labeled a rough guess. You can edit any result before it saves.
The full breakdown of our own run, photo by photo, is on the scan accuracy page, the case for the top spot is on the smartest health scanner, measured, and the scanner guide explains how to get better reads.