Histological image classification: How AI Models Adapt to Unfamiliar Tissue Samples
Source PublicationMicroscopy Research and Technique
Primary AuthorsCelen, Oltulu, Caglar et al.
"Identifying tissues is like training a student to recognise types of forests using only photos of British woodlands. If you then show them photos of the Amazon rainforest, they might struggle. This study tested which AI 'student' could best recognise tissues they had never seen before, proving that some models are much better at adapting to new environments."
The Problem with Histological image classification
Doctors rely on microscopes to diagnose diseases. They examine extremely thin slices of tissue to spot cellular abnormalities. This manual process is slow and requires years of training. Artificial intelligence offers a practical solution. AI can scan these microscopic slides in milliseconds. But there is a major catch. AI models often fail when they look at samples from different hospitals or entirely different species. Histological image classification must be reliable across all types of data to be safe for patients. If an AI learns only on rat tissues, can it accurately identify human tissues? Researchers needed to find out which AI models actually adapt to new information, rather than just memorising their initial training data.
These results were observed under controlled laboratory conditions, so real-world performance may differ.
Testing the Top AI Contenders
Scientists set up a strict benchmarking test. They trained nine different deep neural networks on an internal dataset. This training set contained 4,307 images of rat lung, brain, and fat tissues. These samples were stained with haematoxylin and eosin to give the cells distinct colour. During this initial phase, every single model performed perfectly. They easily learned the local patterns. Then came the real challenge. The team tested the models on a completely new, external set of 600 images. This external dataset contained a diverse mix of human and animal tissues. This out-of-distribution test separates the truly adaptable models from the ones that simply memorised the local rat data.
How the Models Handled Unfamiliar Data
When faced with unfamiliar images, the performance of the models diverged wildly. Simple tissues like fat and brain matter were relatively easy to identify. Most of the competitive models scored over 95 per cent accuracy on these specific tissues. Lung tissue proved much harder to classify. Lungs have a highly varied physical structure. Here, a massive pathology model named UNI2-h took the lead. It achieved a 97.4 per cent accuracy score on the difficult lung samples. It was the absolute best at raw accuracy.
However, accuracy is not the only metric that matters in the real world. Computers in standard medical clinics might not have the power to run massive AI models. The researchers also measured processing speed. The MobileNetV3-Small model was the fastest. It processed images in just 9.4 milliseconds on a basic computer processor. But it lacked the top-tier accuracy of the bigger models.
Impact for Future Clinics
The study suggests a clear path forward for automated medical diagnostics. The massive UNI2-h model is the heavy hitter. It may be the best choice for large research centres with powerful, expensive computers. But the researchers identified a highly effective sweet spot. The ConvNeXt-Small model provided the best overall balance of both metrics. It was highly accurate but still fast enough to run on standard hospital equipment.
These findings could directly shape how healthcare systems build their computer networks. By choosing the right model size, clinics can set up automated support systems that are both fast and reliable. This means doctors could get accurate second opinions almost instantly, potentially speeding up treatments and improving patient care.