Abstract
Objectives
Articular cartilage damage is a primary outcome of pre-clinical and clinical studies evaluating meniscal and cartilage repair or replacement techniques. Recent studies have quantitatively characterized India Ink stained cartilage damage through light reflectance and the application of local or global thresholds. We develop a method for the quantitative characterisation of inked cartilage damage with improved generalisation capability, and compare its performance to the threshold-based baseline approach against gold standard labels.
Methods
The Trainable WEKA Segmentation (TWS) tool (Arganda-Carreras et al., 2017) available in Fiji (Rueden et al., 2017) was used to train two separate Random Forest classifiers to automatically segment cartilage damage on ink stained cadaveric ovine stifle joints. Gold standard labels were manually annotated for the training, validation and test datasets for each of the femoral and tibial classifiers. Each dataset included a sample of medial and lateral femoral condyles and tibial plateaus from various stifle joints, selected to ensure no overlap across datasets according to ovine identifier. Training was performed on the training data with the TWS tool using edge, texture and noise reduction filters selected for their suitability and performance. The two trained classifiers were then applied to the validation data to output damage probability maps, on which a threshold value was calibrated. Model predictions on the unseen test set were evaluated against the gold standard labels using the Dice Similarity Coefficient (DSC) – an overlap-based metric, and compared with results for the baseline global threshold approach applied in Fiji as shown in Figures 1 and 2.
Results
Test set results for the global threshold approach against gold standard labels were 45.0% DSC for the femoral condyle and 32.0% DSC for the tibial plateau. Results for the developed TWS classifiers on the same unseen test data were 79.0% and 72.7% DSC, showing absolute gains of 34.0% and 40.7% DSC over the global threshold baseline for the femoral and tibial classifiers. The trained TWS classifiers were then applied to an external set of unlabelled images of ink stained femoral condyles and tibial plateaus. Model results on sample images shown in Figure 3 further highlight the generalisation capability of the developed models. The most prominent classification features were Hessian filters (32.9%), Entropy (19.4%), Gaussian blur (10.1%), Gabor filters (6.3%) and Sobel filters (6.0%), with all other features contributing less than 6%.
Conclusions
Our findings show that the developed segmentation method more accurately quantifies cartilage damage and provides improved generalisation capability over a range of input variations such as inconsistent orientation and lighting conditions. The developed model enables the use of articular cartilage damage as a reliable and quantitative outcome measure in studies involving large datasets, with reduced requirements for complex pre-processing and specialised equipment.
For any figures or tables, please contact the authors directly.
