littlEye AI 4.0 — a new look and simpler work on Mac, iPad and iPhone. Shared .le3AI and .le3set files remain fully compatible with version 3.x in both directions. CLEARER NAVIGATION • The sidebar has 15 items in the order you work: Overview · Evaluate image · Work (Dataset, Cases, Review, Training, Models) · Research (Quantification, Desmin IHC, Measurement log) · Knowledge (Atlas, Education, Methodology, References) · About. • New Dataset section: all 27 organs as a grid grouped by organ system, with image and labelled counts. • On iPhone, five tabs: Overview · Evaluate · Dataset · Atlas · More. • Overview suggests the next step: import → assign diagnoses → review → train → evaluate and share. Collaboration and support reminders no longer sit above the working sections. • An image in progress, a result or a training run is not lost when you switch sections; training continues and the sidebar shows its progress. RESEARCH TOOLS AS AN INSTRUMENT • Quantification and Desmin IHC: the image and a large result with its interval side by side, settings and verification (fields, point counting, sensitivity) below; study mode as a status strip. • Desmin updates live, without an Evaluate button. Index 2 has a single path — the DAB + H&E image pair; Quantification sends the image to Desmin, not the numbers. • The measurement engine is 5 to 10 times faster and the threshold slider responds instantly. • The Measurement log is its own section: on the Mac a sortable table, deletion of single entries and CSV export via Save As. DATA AND SHARING • Complete backup including the measurement log and staining batches, and a new restore from backup — in Settings. • Reviews in .le3set files are matched through the diagnosis catalogue, images are not re-encoded on import (no quality loss) and exporting a model keeps its original author. • Slide review: digits 1–9, Space skips, ⌘Z undoes the last vote; vote ties are resolved deterministically. • The test set is split by case so that the test result is not inflated. • A damaged archive index is no longer overwritten — the app offers to restore the previous version. MAC • Settings (⌘,) with four tabs: General, Appearance, Reminders, Backup; a native About window. • New shortcuts: ⇧⌘A Dataset, ⇧⌘H Review, ⇧⌘L Measurement log, ⇧⌘I import images, ⇧⌘N new case, ⇧⌘E export model, ⌥⌘A Atlas in a new window; ⌘1–⌘0 stay. • Context menus and an inspector in Models. DESIGN AND ACCESSIBILITY • A unified design derived from the app icon, a full dark mode and colours with verified contrast. • VoiceOver labels for controls and charts, Dynamic Type, numbers formatted for the app language. 447 automated tests.

Quantification methodology to the standard of the scientific community — what so far needed ImageJ or QuPath is now in the app: • Automatic threshold (Otsu): the optical-density histogram of the tissue sets the threshold, not the slider — rater-independent; the value used is in the result and in the Methods paragraph. • Distribution across fields: a scan or large photograph of the section is divided into a grid of fields, each measured with the same recipe and a fixed threshold (systematic sampling). Mean ± SD, coefficient of variation, range and a map (red = low) — for desmin it decides between focal and diffuse loss. Also in the DAB + H&E image pair. • Point counting — stereological check: a grid of ~250 points with a random offset, reclassify points by tapping; the point fraction with a Wilson 95 % CI and a comparison with the pixel value. The reference method against which the automatic figure is verified. • References extended with methods and statistics (Otsu, IHC Profiler, QuPath, Macenko, Gundersen & Jensen, Wilson, Taylor & Levenson, Rimm, Shrout & Fleiss, Koo & Li, Bland & Altman) — on every recipe. The Methods paragraph, the log and the CSV include the new values. Collaboration and validation: a new sheet with a ready multi-centre validation-study proposal (copyable into an e-mail or grant application), the collaboration offer and project support — from the collaboration card and the Help menu. Manual 3.4 (SK/EN). 105 automated tests.

Desmin IHC: the pair of images of the same block (DAB + H&E) now sits directly on the screen — load a scan or microphotograph of the DAB-stained section and a photo of the same block stained with H&E, and you see the CPA, the myocyte area and Index 2 at once. The app checks whether the pair is comparable (format, size, tissue share, nucleus size). Tools for publication — what a reviewer asks about first: • Image quality control: sharpness (the Crete 2007 blur metric), tissue share, clipped highlights and shadows, white point and colour cast — every image gets a badge suitable / borderline / unsuitable with the reason. • Threshold sensitivity analysis: scans the whole slider range, draws the curve, shows the plateau and states whether the value is robust to the threshold; for Index 2 both thresholds are combined into one range. • Stain vectors from your own slide (Macenko 2009) instead of the standard ones, with the deviation from the standard in degrees and anchored units. • Methods paragraph: ready Materials-and-Methods text (SK/EN) with the version, citations, optical-density thresholds, scale, controls and a settings checksum — the same checksum means the same setting. • Log: ICC(2,1) with a confidence interval and a Bland–Altman plot between and within raters; staining batches with a control slide, drift and a normalisation factor; CSV with the checksum and the batch. New manual 3.3 (SK/EN). 97 automated tests including calibration on real myocardium images.

Training without a manual. A dataset readiness traffic light tells you how many images each class is missing — and where to get them; three training profiles (Quick, Standard, Thorough) instead of parameters; the model name fills itself in. After training, one sentence about quality: usable for triage, indicative, unsure about a specific class, or overfitted — with “What to add to improve it” and “Try on an image”. The app reminds you to retrain when new images arrive. Import a model trained outside the app (Core ML .mlpackage / .mlmodel) — with mandatory data provenance and an in-app test on your own images. Mac: Esc closes sheets, resizable Settings, “Open in a separate window” on right-click, training progress in the Dock. Images can also be dropped into Classify image. Fixes: validation no longer shows 0 % on small datasets; bulk image import runs in the background with a counter; two missing links in the Autolysis guide; a false thank-you on launch. New manual 3.1 (SK/EN) in the Help menu and on the website.

From this version littlEye AI is free. No subscription, no feature behind a payment. The app was built for medical students and practising physicians, and the price stood in the way of exactly those people. Inside you will find a voluntary contribution that unlocks nothing. New on Mac: • Menu bar commands with shortcuts — import ⌘O, section switching ⌘1 to ⌘7 • Dropping a .le3AI or .le3set file onto the window imports it • Images can be dragged straight into an organ, on iPad too • References & sources: every diagnosis, stain and IHC marker now carries citations with links to the source

We have clarified what the app is and the information about models. littlEye AI is a tool for training your own models. It ships with a single sample model trained on synthetic patterns — real diagnostic models are ones you train on your own slides, or exchange with colleagues as separate .le3AI files. They never have been and never will be part of the app. We no longer offer models derived from material under a non-commercial licence, and no longer list them in the app. The licence-clean colorectal cancer model (NCT-CRC dataset, CC BY 4.0) remains available free on request at forensika.eu@icloud.com. No functional changes.

A major update to datasets, models and evaluation. RECOGNITION & DIAGNOSTICS • Image evaluation scans overview photos at multiple scales by itself — markedly better results for photos that do not match the training scale • Field-of-view selection with preset squares (½ and ⅓) for a uniform field on every image; new 90° rotation • Low-resolution warning with a concrete recommendation • Model chain: the app recognizes the tissue first, then automatically runs that organ's diagnostic models DATASET & LEARNING • 27 organs and tissues, an atlas of 100+ diagnoses including tumours of the brain, lung, liver, kidney, pancreas, breast, thyroid and melanoma • One-step bulk labelling of an organ's unlabelled images • Multi-file model import; duplicate-proof dataset import — the same image is never added twice • After training, a clear note on where the model is stored, plus direct .le3AI file export RELIABILITY • .le3AI and .le3set files are encrypted and carry a full audit trail including data source and licence • Cloud imports reworked — files are safely downloaded first • Data protection: the store is never overwritten on failure and the app keeps a backup A research and decision-support tool — the final assessment always belongs to the pathologist or forensic physician.

littlEye AI 2.x brings: • A stain field for every image — H&E, PAS, trichrome (Masson), silver, elastic, Luxol, Giemsa and more; the trainer warns about mixed stains • Extended microscope magnifications: from 4× to 1000× • Stain metadata travels in shared datasets (.le3set) and can be edited on any image • A new methodology chapter: why every class should come from at least three different cases — so the model learns tissue, not a slide's hue • A models overview with data provenance in the About screen • Encrypted transfer files — .le3AI and .le3set open only in littlEye AI, and a modified or damaged file is rejected. It happens automatically: no passwords, nothing extra to set up • Learning from corrections — after classifying an image you can correct the diagnosis and add the image to your dataset in one tap, so the next training round learns from it • Fixed importing of files stored in iCloud • On first launch the language follows your device (Slovak, otherwise English) • A welcome screen thanking you for supporting the development