Ensemble learning for robust knee cartilage segmentation: data from the osteoarthritis initiative.

PurposeTo evaluate the performance of an ensemble learning approach for fully automated cartilage segmentation on knee magnetic resonance images of patients with osteoarthritis. Materials and MethodsThis retrospective study of 88 participants with knee osteoarthritis involved the study of three-dimensional (3D) double echo steady state (DESS) MR imaging volumes with manual segmentations for 6 different compartments of cartilage (Data available from the Osteoarthritis Initiative). We propose ensemble learning to boost the sensitivity of our deep learning method by combining predictions from two models, a U-Net for the segmentation of two labels (cartilage vs background) and a multi-label U-Net for specific cartilage compartments. Segmentation accuracy is evaluated using Dice coefficient, while volumetric measures and Bland Altman plots provide complimentary information when assessing segmentation results. ResultsOur model showed excellent accuracy for all 6 cartilage locations: femoral 0.88, medial tibial 0.84, lateral tibial 0.88, patellar 0.85, medial meniscal 0.85 and lateral meniscal 0.90. The average volume correlation was 0.988, overestimating volume by 9% {+/-} 14% over all compartments. Simple post processing creates a single 3D connected component per compartment resulting in higher anatomical face validity. ConclusionOur model produces automated segmentation with high Dice coefficients when compared to expert manual annotations and leads to the recovery of missing labels in the manual annotations, while also creating smoother, more realistic boundaries avoiding slice discontinuity artifacts present in the manual annotations. Key ResultsO_LICombining a 2-label U-Net (cartilage vs background) with a multi-class U-Net for segmentation of cartilage compartment boosts the accuracy of our deep learning model leading to the recovery of missing annotations in the manual dataset. C_LIO_LIAutomatically generated segmentations have high Dice coefficients (0.85-0.90) and reduce inter-slice discontinuity artefact caused by slice wise delineation. C_LIO_LIModel refinement yields more anatomically plausible segmentations where each cartilage label is composed of only a single 3D region of interest. C_LI