Application of Deep Learning Image Segmentation to Synchrotron Radiation µCT Bone Microstructure Datasets

Advancements in 3D and in vivo 4D imaging techniques have allowed for significant developments in bone biological research. High-resolution modalities permit the detection and analysis of features associated with the lacuno-canalicular network (LCN), a framework for bone composition, quality, and strength [1]. The LCN has been the target of pharmaceutical agents for treating bone disorders [2] and is studied in biological anthropology to discern life history information, assess bone quality, and aid in age-at-death estimation. Further investigations of the LCN will inform understandings of bone remodeling and related processes under healthy and pathological conditions (e.g., osteoporosis). Image segmentation of biological data is used to isolate and quantify microstructural parameters of interest. Current techniques include manual segmentation and the application of thresholding and morphological operations. The former is time-consuming, and both are subject to error. The employment of deep learning segmentation, however, allows for consistent semi-automatic segmentation of data and has the potential to improve processing times and accuracy. Previous research [3] by our group indicates that deep learning image segmentation can achieve comparable results to established protocols for bone microstructure segmentation. The primary objective of this work is to improve upon existing results by using separate deep learning models for vascular pores and lacunae, while utilizing morphological operations to refine deep learning segmentation results. Healthy and pathological human left sixth rib samples from decedents of varying ages (21-54 years-at-death, mean age = 38 years) were obtained through collaboration with an American non-profit organ procurement organization. Cylindrical cores of cortical bone were procured for imaging experiments. Synchrotron µCT data were obtained using the BMIT Beamlines at the Canadian Light Source facility. The setup was equipped with a white beam microscope and a 5x objective lens to achieve a resolution of 1.5 µm. Four training datasets of ∼100 slices each were prepared through manual segmentation, thresholding, and morphological operations and reviewed to correct any errors. The most promising architecture was selected by training and testing models based on several architectures available in ORS Dragonfly [4]. Two models based on the same architecture were trained, one for the segmentation of vascular pores and another for lacunae. Results were compared against an established bone microarchitecture segmentation method using CTAN (Bruker). Manual segmentation and deep learning model training were performed in ORS Dragonfly [4]. The CTAN method relies exclusively on global thresholding and morphological operations. The pixel intensity thresholds are decided, and several automatic morphological operations are applied to refine the thresholding segmentation, including removing noise and closing gaps. Vascular pores and lacunae from ten comparative samples, distinct from the training datasets, were segmented to compare the two segmentation methods; the established protocol in CTAN and deep learning segmentation. Also, the data were manually segmented using the same method as the training datasets, and then used as a reference to compare the accuracy of the two methods. Descriptive statistics were extracted from the data and Levene’s and Shapiro-Wilk normality tests were conducted. If the samples violated normality, they were bootstrapped. Several parameters, including total lacunar and pore volumes, and average lacunar and pore sizes, were extracted from the segmentations and compared using t-tests and one-way ANOVAs. The segmentation results of the two tested protocols were also compared based on the Dice Similarity Coefficient (DSC), Accuracy, and True Positive and Negative rates. Previous work [3] demonstrated that deep learning segmentation of bone microarchitectural data provides comparable results to an established protocol (CTAN). Our preliminary work using the UNet++ deep learning architecture [5] has indicated that using two deep learning models to segment pores and lacunae separately (two classes per model, where one class is the relevant structure and the other the background) can provide more accurate segmentations than a three-classes model segmenting pores and lacunae simultaneously. Relevant DSC values were obtained using a subsection of the training dataset as reference to compare the trained models' effectiveness. The three-classes lacunae and pores model achieved a DSC of 0.9810, while the two-classes models achieved a DSC of 0.9864 for lacunae and 0.9959 for pores. The improvement in DSC, and thus segmentation accuracy, when using two separate segmentation models indicates a potential that is worthy of more investigation. Additionally, morphological operations can be used to refine segmentations by eliminating noise and closing segmentation gaps, as seen in the established CTAN protocol. Such operations can be incorporated into a deep learning-based bone segmentation workflow. Combined with the improved accuracy of employing individual models for defined parameters, these changes to the segmentation protocol have the potential to increase the accuracy and precision of feature definition in bone, increase segmentation consistency, and reduce observer error.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.