Forensika › Apps › littlEye AI
Teach your own AI model to recognise histology slides directly on an iPhone, iPad or Mac: with no internet and without a single pixel leaving the device. Free since version 3.0.
Most AI tools for histology image recognition send images to a server: for forensic practice working with images of real autopsy specimens, that is a problem in its own right, independent of model accuracy. littlEye AI takes the opposite route: both training and evaluation run through Core ML directly on an iPhone, iPad or Mac, an image never leaves the device, and the app has no networking layer and no telemetry.
The H&E slide dataset spans 27 organs and tissues; slides are reviewed by colleagues with an audit trail, and the final diagnosis is settled by majority vote. Training always sets aside an independent test set and reports a confusion matrix with per-class sensitivity and precision: students and physicians alike see exactly where a model fails, not only where it succeeds.
A colour-coded rating of the dataset shows whether you have enough varied images per class (30 or more) before you start training.
.le3AI models and .le3set datasets are AES-GCM encrypted and carry their own metrics and audit trail: shared between colleagues via AirDrop or email.
A knowledge base with microscopic descriptions, key features and suggested conclusions, drawn from cited literature.
Area, count and size of colour-defined structures measured on the image, a desmin IHC calculator and tools for publication: image quality control, threshold sensitivity, ICC and Bland–Altman.
A misread slide is added back to the dataset in one tap, so the next training round learns from it.
Microphotographs (JPEG/PNG/TIFF/HEIC) are sorted by organ and diagnosis; a stain and magnification log keeps the data consistent.
Multiple colleagues independently review slides; the app records the reviewer's name and date for each assessment (audit trail), and majority vote sets the final label.
Configurable iterations and augmentations, an independent test set (10–20% stratified holdout), progress with the option to cancel.
Training, validation and test accuracy, plus sensitivity and precision reported separately for each class.
The finished model is saved as an encrypted .le3AI file with a manifest: imported into another littlEye AI installation via Files, AirDrop or a double tap.
littlEye AI is not only a training tool: it also offers a full case workflow: scanner slides attached to a case number, case history and the post-mortem conclusion as inputs, a tissue check that flags a possible slide mix-up, and a concise summary of findings copied to the clipboard in one tap. An education module adds a "what next after H&E" guide, special stains, IHC markers, cellular elements and pigments: each entry carrying a source citation.
| Platform | iPhone, iPad (iOS/iPadOS 17+), Mac (macOS 14+) |
| Dataset | 27 organs and tissues, your own images |
| Atlas | 103 diagnoses with microscopic descriptions and citations |
| Formats | .le3AI (model), .le3set (dataset) · encrypted |
| Price | free since version 3.0 (donationware) |
No. The app is a training tool, not a model collection: it ships with a single demonstration model trained on synthetic patterns. Real models are trained from your own slides, or exchanged with colleagues as standalone encrypted .le3AI files.
No. The entire app runs exclusively on-device: images, cases and trained models never leave it unless you explicitly export them. The app makes no network connections and has no telemetry or analytics.
The technical minimum is 2 classes with 5 images each, but the app recommends 30 or more images per class drawn from multiple cases: a colour-coded dataset-readiness indicator shows whether you have enough material before you start training.
No. It is a research and decision-support tool for qualified professionals. It does not provide a diagnosis: the final assessment always belongs to the pathologist or forensic physician.
Yes, since version 3.0. No subscription, no feature behind a paywall: an optional one-time in-app contribution supports further development and unlocks nothing.