Automatic knee meniscus tear detection and orientation classification with Mask-RCNN.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Orthopedics

Background:

• Knee meniscus tears are common injuries requiring accurate diagnosis.
• Magnetic Resonance Imaging (MRI) is a key modality for evaluating knee injuries.
• Automated analysis of medical images can aid radiologists in diagnosis.

Purpose of the Study:

• To classify knee MRI images for the presence, location, and orientation of meniscal tears.
• To contribute to a data challenge focused on meniscus tear classification.
• To develop a robust deep learning model for meniscus tear analysis.

Main Methods:

• A mask region-based convolutional neural network (R-CNN) was trained for meniscus localization.
• Ensemble aggregation was used to enhance the robustness of the R-CNN model.
• A shallow Convolutional Neural Network (ConvNet) was cascaded for tear orientation classification.

Main Results:

• The developed approach accurately predicted meniscal tears in the challenge dataset.
• The strategy achieved a weighted AUC score of 0.906 across all three classification tasks.
• The method ranked first in the French Radiology Society's data challenge.

Conclusions:

• The model demonstrated high performance in classifying knee meniscus tears.
• Future improvements could involve expanding the dataset or utilizing 3D MRI data.
• Enhanced performance is expected for complex cases like extensive or multiple meniscal tears.