Automating three-dimensional osteoarthritis histopathological grading of human osteochondral tissue using machine learning on contrast-enhanced micro-computed tomography
Rytky, Santeri (2019-05-13)
Rytky, Santeri
S. Rytky
13.05.2019
© 2019 Santeri Rytky. Tämä Kohde on tekijänoikeuden ja/tai lähioikeuksien suojaama. Voit käyttää Kohdetta käyttöösi sovellettavan tekijänoikeutta ja lähioikeuksia koskevan lainsäädännön sallimilla tavoilla. Muunlaista käyttöä varten tarvitset oikeudenhaltijoiden luvan.
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:oulu-201905141769
https://urn.fi/URN:NBN:fi:oulu-201905141769
Tiivistelmä
Osteoarthritis (OA) is a joint disease affecting hundreds of millions of people worldwide. In basic research, accurate ex vivo measures are needed for assessing OA severity. The standard method for this is the histopathological grading of stained thin tissue sections. However, the methods are destructive, time-consuming, do not describe the full sample volume and provide subjective results. Contrast-enhanced micro-computed tomography (CEμCT) -based grading with phosphotungstic acid -stain was previously developed to address some of these issues. Aim of this study was to investigate the possibility of automating this process.
Osteochondral tissue cores were harvested from total knee arthroplasty patients (n = 34, N = 19, Ø = 2 mm, n = 15, N = 5, Ø = 4 mm) and asymptomatic cadavers (n = 30, N = 2, Ø = 4 mm). Samples were imaged with CEμCT, reconstructed and graded manually. Subsequently, the reconstructions were loaded into an ad hoc developed Python software, where volumes-of-interest (VOI) were extracted from different cartilage zones: surface zone (SZ), deep zone (DZ) and calcified zone (CZ) and collapsed into two-dimensional texture images.
Normalized images underwent Median Robust Extended Local Binary Pattern (MRELBP) -algorithm to extract the features, with subsequent dimensionality reduction. Ridge and logistic regression models were trained with L2 regularization against the ground truth for the small samples (Ø = 2 mm) using leave-one-patient-out cross-validation. Trained models were then evaluated on the large samples (Ø = 4 mm). Performance of the models were assessed using Spearman’s correlation, Area under the Receiver Operating Characteristic Curve (AUC) and Average Precision (AP).
Highest performance on both models was for the SZ. Strong correlation was observed on ridge regression (ρ = 0.68, p < 0.0001), as well as high AUC and AP values for the logistic regression (AUC = 0.92, AP = 0.89) for the small samples. Using the large samples, similar findings were observed with slightly reduced values (ρ = 0.55, p = 0.0001, AUC = 0.86, AP = 0.89). Moderate results were observed for CZ and DZ models (ρ = 0.54 and 0.38, AUC = 0.77 and 0.72, AP = 0.71 and 0.50, respectively). Evaluation on the large samples resulted in performance decrease on CZ models (ρ = 0.29, AUC = 0.63, AP = 0.62), while surprisingly performance increased on DZ logistic regression model (ρ = 0.34, AUC = 0.72, AP = 0.83).
Obtained results indicate that automating the 3D CEμCT histopathological grading is feasible. However, with low number of samples, models are better suited for binary detection of sample degenerative features, rather than predicting a detailed grade. To facilitate model generalization on new data, similar data acquisition protocol should be used on all samples. The proposed methods have potential to aid OA researchers and pathologists in 3D histopathological grading, introducing more objectivity to the grading process. This thesis presents the conducted study in detail, and provides an extensive review related to the osteochondral unit, CEμCT imaging, as well as statistical learning machines.
Osteochondral tissue cores were harvested from total knee arthroplasty patients (n = 34, N = 19, Ø = 2 mm, n = 15, N = 5, Ø = 4 mm) and asymptomatic cadavers (n = 30, N = 2, Ø = 4 mm). Samples were imaged with CEμCT, reconstructed and graded manually. Subsequently, the reconstructions were loaded into an ad hoc developed Python software, where volumes-of-interest (VOI) were extracted from different cartilage zones: surface zone (SZ), deep zone (DZ) and calcified zone (CZ) and collapsed into two-dimensional texture images.
Normalized images underwent Median Robust Extended Local Binary Pattern (MRELBP) -algorithm to extract the features, with subsequent dimensionality reduction. Ridge and logistic regression models were trained with L2 regularization against the ground truth for the small samples (Ø = 2 mm) using leave-one-patient-out cross-validation. Trained models were then evaluated on the large samples (Ø = 4 mm). Performance of the models were assessed using Spearman’s correlation, Area under the Receiver Operating Characteristic Curve (AUC) and Average Precision (AP).
Highest performance on both models was for the SZ. Strong correlation was observed on ridge regression (ρ = 0.68, p < 0.0001), as well as high AUC and AP values for the logistic regression (AUC = 0.92, AP = 0.89) for the small samples. Using the large samples, similar findings were observed with slightly reduced values (ρ = 0.55, p = 0.0001, AUC = 0.86, AP = 0.89). Moderate results were observed for CZ and DZ models (ρ = 0.54 and 0.38, AUC = 0.77 and 0.72, AP = 0.71 and 0.50, respectively). Evaluation on the large samples resulted in performance decrease on CZ models (ρ = 0.29, AUC = 0.63, AP = 0.62), while surprisingly performance increased on DZ logistic regression model (ρ = 0.34, AUC = 0.72, AP = 0.83).
Obtained results indicate that automating the 3D CEμCT histopathological grading is feasible. However, with low number of samples, models are better suited for binary detection of sample degenerative features, rather than predicting a detailed grade. To facilitate model generalization on new data, similar data acquisition protocol should be used on all samples. The proposed methods have potential to aid OA researchers and pathologists in 3D histopathological grading, introducing more objectivity to the grading process. This thesis presents the conducted study in detail, and provides an extensive review related to the osteochondral unit, CEμCT imaging, as well as statistical learning machines.
Kokoelmat
- Avoin saatavuus [34516]