Agreement and accuracy of fully automated morphometric femorotibial cartilage analysis in radiographic knee osteoarthritis

Objective

To examine the performance of automated convolutional neuronal network cartilage segmentation in knees with radiographic osteoarthritis (ROA), and its dependence on the OA status of the training set and MRI sequences.

Design

We studied 122 ROA and 92 healthy reference cohort (HRC) knees from the Osteoarthritis Initiative. All knees had expert manual segmentation of the femorotibial cartilages based on coronal FLASH and sagittal DESS MRI. Two U-net convolutional neural networks were trained on 86/50 ROA/HRC knees, validated on 18/21, and tested on 18/21.

Results

Of 122 ROA knees, 43 (35%) were KLG2, 41 (34%) KLG3, and 38 (31%) KLG4. In the ROA test set, Dice Similarity Coefficients (DSCs) were 0.86/0.86 (FLASH/DESS) for the algorithm trained on ROA, and 0.82/0.82 for that trained on HRC knees. In the HRC test set, mean DSCs of 0.91/0.90 were observed with FLASH/DESS, both with the HRC- and with the ROA-trained algorithm. Cartilage thickness computation from automated segmentation in the FLASH ROA test set showed a correlation of r = 0.94 with manual segmentation for the ROA-trained, and of r = 0.89 for the HRC-trained algorithm. In the FLASH HRC test set, it exhibited a correlation of r = 0.96 for the HRC-trained and r = 0.88 for the ROA-trained algorithm. Results were similar between the DESS and FLASH, but were less accurate for KLG4 than for KLG2/3 knees.

Conclusions

An automated algorithm trained on ROA knees was able to accurately segment and compute cartilage thickness on FLASH and DESS MRI, in both ROA and healthy knees, whereas an algorithm trained on healthy knees did not perform as well in ROA knees.