Fully and Weakly Supervised Deep Learning for Meniscal Injury Classification, and Location Based on MRI.

Area of Science:

• Orthopedics
• Radiology
• Artificial Intelligence

Background:

• Meniscal injuries are a frequent cause of knee pain and a precursor to knee osteoarthritis (KOA).
• Accurate segmentation and classification of meniscal injuries are crucial for effective diagnosis and treatment planning.
• Current diagnostic methods can be time-consuming and subject to inter-observer variability.

Purpose of the Study:

• To develop an automated pipeline for meniscal injury classification and localization using deep learning (DL) models.
• To evaluate the performance of fully and weakly supervised DL networks on MRI images for meniscus segmentation and injury classification.
• To compare the diagnostic performance of the developed DL models against human radiologists.

Main Methods:

• Development of the LGSA-UNet model utilizing feature fusion from adjacent MRI slices for enhanced contextual information.
• Training segmentation and classification models on a large dataset of 1,756 knees from the Osteoarthritis Initiative (OAI).
• External validation of the models using 206 knees from an orthopedic hospital and comparison with junior and senior radiologists' diagnoses.

Main Results:

• The segmentation model achieved a DICE coefficient between 0.84 and 0.93.
• Binary classification models demonstrated Area Under the Curve (AUC) values ranging from 0.85 to 0.95.
• Classification accuracy for normal, tear, and maceration types ranged from 0.60 to 0.88, with the DL model outperforming a junior radiologist.

Conclusions:

• The developed DL pipeline shows significant potential for accurate knee meniscus segmentation and injury classification.
• The automated approach can improve diagnostic efficiency and consistency in identifying meniscal pathologies.
• This technology could aid radiologists in clinical practice, leading to better patient outcomes for knee osteoarthritis and related conditions.