Deep learning-based prediction of knee osteoarthritis severity score using histopathological images of bone-cartilage specimens and YOLOv8

Abstract

This study aims to classify the histopathological HE- and SafO-stained images of bone-cartilage specimens automatically, using YOLOv8 by severity.

Materials and methods. The pre-trained YOLOv8m-cls neural network of the Ultralytics package was used. Python 3.9. was used. Models were trained on a Human Knee Cartilage Histopathology Assessment Dataset.

Results. Machine learning models were built to score the severity levels of knee oasteoarthritis using the histopathological images of the bone-cartilage specimens. The accuracy_top1 performance was 95.7% for the model trained on HE-images and 94.3% - on SafO images.

Conclusion. YOLOv8 can be used in histopathology with high performance. Models can be used by researchers to automate the histopathology severity scoring of knee osteoarthritis. The dataset upscaling and the model modifying could help to improve the quality of predictions.