Written by Liam Cattell, Computer Vision Lead
Artificial intelligence may sound like an all-encompassing phrase evoking thoughts of robot dogs, sci-fi androids, and chat bots. However, when discussing artificial intelligence within the realm of veterinary pathology the aim is clear: make pathology more accurate and repeatable, and build a library of reference intervals with unprecedented potential.
Moichor’s AI resembles the learning process budding human pathologists might undertake in their studies and professional careers. Essentially, the Moichor AI works by learning, making mistakes during training, and improving like a human; it absorbs large quantities of high-quality data, attempts to identify cells in a blood smear, and updates its knowledge of cell characteristics by comparing its performance to our team of pathologists.
When pathologists identify cells in a blood smear, they draw on experience acquired through study and practice. During their years of training, they receive grades and guidance from teachers and colleagues that help them improve their understanding of cell pathology. By learning under the supervision of an expert, pathologists (and their patients) can be confident in their assessments. This learning system is, unsurprisingly, called supervised learning, and it is also one of the fundamental concepts in artificial intelligence.
For Moichor, this means our AI models are trained to identify leukocytes in a supervised manner, just like pathologists. We provide the computer with blood smear images containing thousands of expertly annotated cells.
Learning is an iterative process in which the model constantly compares its predictions with those expert annotations and updates its internal state to improve its performance. Once the training is complete, the model can be used to identify cells in images it has never seen before.
While AI algorithms are trained to identify patterns in data, they do not have the same understanding of the context of the data as pathologists. A pathologist can provide context for our AI models by explaining the meaning of the results and how they relate to the clinical question at hand. Therefore, we employ a “human-in-the-loop” strategy, whereby assessments from our AI models are supervised for a second time and validated by a person before they are returned to veterinarians. By having a person-in-the-loop, we can improve the accuracy, reliability, and confidence of our AI-assisted pathology. This is why we say our AI is supervised twice, and having this additional quality check can lead to better patient care and outcomes.
Although our AI models have probably seen more cells and species than most pathologists (4.3 million leukocytes, 780 species and counting), they can still make mistakes. White blood cells are a diverse group of cells with a wide range of shapes, sizes, and functions. Their appearance can vary depending on the patient’s health, sex, age, and many other factors. This can make it difficult for both AI algorithms and pathologists to identify them accurately. We counter this problem by training our models with a large, diverse set of data. In some situations, we even deliberately degrade the training images’ quality to make the learning process harder and the models more robust.
Another challenge facing AI models is the prevalence of different cell types. Pathologists and AI algorithms can identify healthy neutrophils with relative ease due their abundance in blood smears, but rare blast cells and metamyelocytes can confound both human and machine. For this reason, our models work in tandem with our team of pathologists, combining the expertise from both camps to deliver reliable results. Furthermore, we are constantly monitoring our mistakes and using them to guide future development of improved AI models.
At Moichor, we often encounter new species or previously unseen health issues that push the limits of our AI models. As the diversity of our blood smear database grows, we incorporate new features into our models to reflect the current reality of the data we receive. For example, if we face an increased number of blood smears from sick bearded dragons, we probably want our model to improve its detection of toxic heterophils. This approach helps combat “model drift”, where models that were once accurate may become less accurate over time due to the changing nature of real-world data. We regularly refresh our models with new training data to keep them up-to-date and improve their accuracy, which is part of our larger mission of building an extensive reference interval library.
New data can also help to prevent bias. Model bias is a problem that can occur when a model is trained on data that is not representative of the population that the model is intended to be used on. For example, a model trained to identify leukocytes in parrot blood smears would be less suitable for assessing blood smears from chickens. Even a model trained on parrots may become less accurate as the number of parrot species increases. By expanding the diversity of the training dataset, we are more likely to capture a wider range of health conditions and image variation. This helps to ensure that the model is not biased towards any particular type of blood smear.
Not only do we update our training datasets, but we also update the AI models themselves. AI research is progressing at an incredible rate, with faster, fairer, and more accurate models being reported in the literature every few weeks. We track trends in both veterinary and non-veterinary AI research and apply the best models for our needs, even if we have to adapt them from other AI domains.
Moichor’s AI is growing under the watchful eyes of veterinary pathologists and data scientists, and as the technology develops, our team sees many opportunities to improve animal health across the spectrum.
Whether it’s helping pathologists to provide a deeper level of insight for each slide or making pathology review on every sample possible, Moichor will continue to diligently train our artificial intelligence using rigorous methodology and the expertise of human laboratory staff, data scientists, and veterinary pathologists.