Diagnosing orthopaedic infection by identifying neutrophils in whole histology slide images with machine learning trained on publicly available datasets

Listen to Simon and Amy discuss the paper 'Diagnosing orthopaedic infection by identifying neutrophils in whole histology slide images with machine learning trained on publicly available datasets' published in the March 2026 issue of Bone & Joint Research.

Click here to read the paper.

Be the first to know when the next episode is live! Follow our social media accounts, @BoneJointPortal and @BoneJointRes on X, and @bonejointres on Instagram for updates!

Transcript

[00:00:00] Welcome back to another episode of AI Talks with Bone & Joint from the publishers of Bone & Joint Research. Today, we're discussing a paper titled 'Diagnosing orthopaedic infection by identifying neutrophils in whole histology slide images with machine learning trained on publicly available datasets', published in March 2026 by K Bentick and colleagues. I'm Simon and I'm Amy.

This study is really interesting as it explores using machine learning to automate a vital part of diagnosing periprosthetic joint infections, or PJIs. PJI is a serious complication that can occur after joint arthroplasty surgeries, affecting about one to 2% of such procedures in the UK.

The paper primarily focuses on the YOLO, or You Only Look Once, object detection model. This cutting-edge method has been trained on publicly available datasets to identify and count neutrophils in histology slides. The goal is to automate a task that is typically done manually, which can be very time-consuming and prone to human error.[00:01:00] 

Exactly and the researchers also wanted to see if this automated approach could match the accuracy and reliability of traditional diagnostic methods. They used a dataset of 3,923 images, which included blood film microscopic slides with neutrophils, supplemented with additional histological slides taken at the time of revision surgery.

They divided these images into training, validation, and test sets to train the YOLO model. Then they tested the model's performance against 19 additional cases which the model had not previously seen. Simon, the results were quite impressive! The model achieved a precision of 82%, a recall of 79%, and an F1 score of 80% on the ground truth images.

When compared with formal histopathological, microbiological, and multidisciplinary team diagnoses, the precision and recall metrics remained comparably high.

Yes, those metrics show that the model can identify neutrophils almost as well as human [00:02:00] experts. They even tested it on more complex tissue samples and found remarkable consistency. For diagnosing infection, the model achieved a recall of 82% and an F1 score of 86% when compared to the MDT diagnosis.

Another significant finding was that the model accurately identified 9 out of 10 infected cases and correctly ruled out 7 out of 9 non-infected cases. This demonstrates its potential to significantly aid pathologists by pinpointing hotspots of neutrophil activity indicative of infection.

Simon, it's exciting to think about how this could be used in clinical settings. Pathologists are often in short supply, and their workload is immense. An automated system like this could streamline the diagnostic process and free up valuable time for pathologists to focus on more complex cases.

Certainly, but the authors did note some limitations. For example, the study was retrospective, which introduces the risk of selection bias. They also mentioned that different [00:03:00] laboratories, tissue processing techniques, and scanners might affect the model's generalizability. Thus, further perspective studies are needed.

That's an important point. Real world applications often reveal challenges not anticipated in controlled settings. For instance, different staining techniques or image resolutions could impact the model's effectiveness, and additional training might be needed to adapt to those variables.

One intriguing aspect was their decision to keep the pathologist involved. The model provides a count of neutrophils, but the final diagnosis is left to a human expert. This allows the pathologist to validate and correct the model's predictions, which can then be used to further refine the model. Lastly, the authors highlighted the potential of using YOLO in real-time applications, like mounting it onto microscopes to help pathologists focus on areas with a high density of neutrophils. This could greatly enhance the efficiency and accuracy of diagnoses.

In summary, this study marks an important step towards integrating machine learning into [00:04:00] pathology. It shows that machine learning models, specifically YOLO, can accurately identify neutrophils and assist in diagnosing orthopedic infections.

It's a promising development that could have a significant impact in clinical practice, improving both efficiency and diagnostic accuracy. Thanks for tuning in to another episode of AI Talks with Bone & Joint. Make sure to check out the full study for a more detailed understanding. Until next time, I'm Simon and I'm Amy. Goodbye and take care.