Forensika › Articles

Histology · artificial intelligence

The model said “heart, 100 %”. It was a normal lung.

Ján Šikuta, MD, PhD, forensic pathologist · Published 15 November 2026

An image of normal lung from the web was labelled heart by the model, with 100 % confidence. Grey hepatisation in lobar pneumonia came out as liver, 81.8 %. While developing littlEye AI, mistakes like these taught me more than good figures did.

Why a model is wrong so confidently

A model cannot say “I don't know”. It has to pick one of its classes, so 100 % does not mean a pathologist's certainty. It only means that, of everything it knows, the image looked most like heart. It had never seen pneumonia in training, and a dense airless lobe looked to it like liver. That is why littlEye AI runs a diagnostic model only on the tissue it belongs to and shows the result as a proposal with a percentage.

What these mistakes taught me

  • Magnification has to match. A model learns the size of cells in pixels. Trained at a 20× objective, it scored 78 % on the test. When we altered the test images to imitate a 40× objective, it fell to 52 %, and to 11 % at 100× oil immersion, which with 19 tissues is close to chance. Images from a microscope therefore need calibrating with a stage micrometer.
  • One slide teaches the shade, not the tissue. With one or two slides per class, the model linked the colour of the preparation to the organ. On foreign material the test fell to about 65 %, and a redder liver from another department came out as kidney. Each diagnosis needs at least three different cases.
  • The test has to be honest. Crops from the same part of a slide do not belong in both training and test, otherwise the model is examined on what it has already seen. Crops also need checking by eye; an automatic filter lets through tissue other than what the label says.
  • Even a freely available collection has a licence. I deleted models from one open collection with a non-commercial licence after the university's statement. Since then I have worked with my own slides and licence-clean data.

This is not just my experience. Tellez et al. (2019) describe, on H&E images from 9 laboratories, how a network trained in one laboratory often performs worse on images from another. Howard et al. (2021) showed in the TCGA database that the site of origin can be recognised from the slide even after colour normalisation, and that this leads to over-optimistic results. With any impressive figure in a paper it is worth asking whether the model was tested on data from another site.

What littlEye AI is and is not

littlEye AI is my app for iPhone, iPad and Mac, free of charge. A physician trains a model from their own H&E micrographs directly on the device; nothing goes to a server. Colleagues can review images, each review stays on record with name and date, and the majority decides. Before training, a traffic light points out weak spots in the data; after training, the app shows the result on an independent test set and a confusion matrix. Departments can exchange sets and models so that each annotates what it knows best.

Validation on foreign material, another scanner and other hands is only prepared so far, so I do not yet have figures on how the models behave outside the department where they were made. That is also why results in the app carry the same sentence: A research and decision-support tool. Not a medical device. The final assessment always belongs to the pathologist or forensic physician.

If you would like to try the models or the quantification on your own material, the terms are in the Collaboration section of the app's website. The study proposal is ready: the app is free, images are not sent, and every participating department is a co-author. I am also looking for H&E and IHC slides with a clear licence to share.

Related app: littlEye AI. This article is a professional overview for qualified experts; it does not replace a textbook, guidelines or an expert opinion.

References

  1. Šikuta J. How littlEye AI learns, and why it sometimes answers wrongly. App guide. https://dzanino.github.io/littlEyeAI/littlEye_AI_how_it_learns_EN.pdf
  2. Tellez D, Litjens G, Bándi P, et al. Quantifying the effects of data augmentation and stain color normalization in convolutional neural networks for computational pathology. Med Image Anal. 2019;58:101544. doi:10.1016/j.media.2019.101544
  3. Howard FM, Dolezal J, Kochanny S, et al. The impact of site-specific digital histology signatures on deep learning model accuracy and bias. Nat Commun. 2021;12(1):4423. doi:10.1038/s41467-021-24698-1