Research on the morphological structure of partial fracture healing process in diabetic mice based on synchrotron radiation phase-contrast imaging computed tomography and deep learning

Abstract

The prevalence of diabetes mellitus has exhibited a notable surge in recent years, thereby augmenting the susceptibility to fractures and impeding the process of fracture healing. The primary objective of this investigation is to employ synchrotron radiation phase-contrast imaging computed tomography (SR-PCI-CT) to examine the morphological and structural attributes of different types of callus in a murine model of diabetic partial fractures. Additionally, a deep learning image segmentation model was utilized to facilitate both qualitative and quantitative analysis of callus during various time intervals. A total of forty male Kunming mice, aged five weeks, were randomly allocated into two groups, each consisting of twenty mice, namely, simple fracture group (SF) and diabetic fracture group (DF). Mice in DF group were intraperitoneally injected 60 mg/kg 1 % streptozotocin(STZ) solution for 5 consecutive days, and the standard for modeling was that the fasting blood glucose level was ≥11.1 mmol /l one week after the last injection of STZ. The right tibias of all mice were observed to have oblique fractures that did not traverse the entire bone. At three, seven, ten and fourteen days after the fracture occurred, the fractured tibias were extracted for SR-PCI-CT imaging and histological analysis. Furthermore, a deep learning image segmentation model was devised to automatically detect, categorize and quantitatively examine different types of callus. Image J software was utilized to measure the grayscale values of different types of callus and perform quantitative analysis. The findings demonstrated that:1)SR-PCI-CT imaging effectively depicted the morphological attributes of different types of callus of fracture healing. The grayscale values of different types of callus were significantly different(P < 0.01).2)In comparison to the SF group, the DF group exhibited a significant reduction in the total amount of callus during the same period (P < 0.01). Additionally, the peak of cartilage callus in the hypertrophic phase was delayed.3)Histology provides the basis for training algorithms for deep learning image segmentation models. The deep-learning image segmentation models achieved accuracies of 0.69, 0.81 and 0.733 for reserve/proliferative cartilage, hypertrophic cartilage and mineralized cartilage, respectively, in the test set. The corresponding Dice values were 0.72, 0.83 and 0.76, respectively. In summary, SR-PCI-CT images are close to the histological level, and a variety of cartilage can be identified on synchrotron radiation CT images compared with histological examination, while artificial intelligence image segmentation model can realize automatic analysis and data generation through deep learning, and further determine the objectivity and accuracy of SR-PCI-CT in identifying various cartilage tissues. Therefore, this imaging technique combined with deep learning image segmentation model can effectively evaluate the effect of diabetes on the morphological and structural changes of callus during fracture healing in mice.

Keywords: Cartilage callus; Deep learning; Diabetes; Incomplete fracture; Synchrotron radiation phase-contrast CT imaging.