Semantic Segmentation of Micro-CT Images to Analyze Bone Ingrowth into Biodegradable Scaffolds.

The healing of bone fractures is a complex and well-orchestrated physiological process, but normal healing is compromised when the fracture is large. These large non-union fractures often require a template with surgical intervention for healing. The standard treatment, autografting, has drawbacks such as donor site pain and limited availability. Biodegradable scaffolds developed using biomaterials such as bioactive glass are a potential solution. Investigation of bone ingrowth into biodegradable scaffolds is an important aspect of their development. Micro-CT (µ-CT) imaging is widely used to evaluate and quantify tissue ingrowth into scaffolds in 3D. Existing segmentation techniques have low accuracy in differentiating bone and scaffold, and need improvements to accurately quantify the bone in-growth into the scaffold using µ-CT scans. This study proposes a novel 3-stage pipeline for better outcome. The first stage of the pipeline is based on a convolutional neural network for the segmentation of the scaffold, bone, and pores from µ-CT images to investigate bone ingrowth. A 3D rigid image registration procedure was employed in the next stage to extract the volume of interest (VOI) for the analysis. In the final stage, algorithms were developed to quantitatively analyze bone ingrowth and scaffold degradation. The best model for segmentation produced a dice similarity coefficient score of 90.1, intersection over union score of 83.9, and pixel accuracy of 93.1 for unseen test data.

Detection and Tracking Volumes of Interest in 3D Printed Tissue Engineering Scaffolds using 4D Imaging Modalities.

Additive manufacturing (AM) platforms allow the production of patient tissue engineering scaffolds with desirable architectures. Although AM platforms offer exceptional control on architecture, post-processing methods such as sintering and freeze-drying often deform the printed scaffold structure. In-situ 4D imaging can be used to analyze changes that occur during post-processing. Visualization and analysis of changes in selected volumes of interests (VOIs) over time are essential to understand the underlining mechanisms of scaffold deformations. Yet, automated detection and tracking of VOIs in the 3D printed scaffold over time using 4D image data is currently an unsolved image processing task. This paper proposes a new image processing technique to segment, detect and track volumes of interest in 3D printed tissue engineering scaffolds. The method is validated using a 4D synchrotron sourced microCT image data captured during the sintering of bioactive glass scaffolds in-situ. The proposed method will contribute to the development of scaffolds with controllable designs and optimum properties for the development of patient-specific scaffolds.

Detection of Diabetes by Macrovascular Tortuosity of Superior Bulbar Conjunctiva.

More than 8% of world population have diabetes which causes long term complications such as retinopathy, neuropathy, nephropathy and foot ulcers. Growing patient numbers has prompted large scale screening methods to detect early symptoms of diabetes (rather than elevated blood glucose levels which is a late symptom). Vascular tortuosity (twisted and curved nature of blood vessels) in retinal fundus images has proven to reflect the effect of diabetes on macrovasculature. However, large scale patient screening using retinal fundus images has limitations due to the requirement of a retinal camera. Therefore, we hypothesize that the vasculature of superior bulbar conjunctiva which could be captured using a regular camera could be used to measure tortuosity instead of retinal fundus images enabling mass screening.To test this hypothesis, a total of 168 scleral images were acquired from 50 healthy subjects and 34 diabetic patients using a digital camera. The sclera region was segmented using Chan-Vese algorithm and macrovasculature of superior bulbar conjunctiva was segmented using B-COSFIRE filters. Results revealed that the superior bulbar conjunctival macrovascular tortuosity of diabetic patients was significantly less than that of non-diabetic group (p-value =0.015). A similar result was yielded (p-value =0.049) from a group of participants who were less than 40 years old which excluded the age related variation of tortuosity.

A non-rigid image registration technique for 3D ultrasound carotid images using a "twisting and bending" model.

Atherosclerosis at the carotid bifurcation resulting in cerebral emboli is a major cause of ischemic stroke. Most strokes associated with carotid atherosclerosis can be prevented by lifestyle/dietary changes and pharmacological treatments if identified early by monitoring carotid plaque changes. Sensitive monitoring of plaque changes in volume and morphology requires that 3D ultrasound (US) images of carotid plaque obtained at different time points be registered and evaluated for change. This registration technique should be non-rigid, since different head positions in image acquisitions cause relative bending and torsion in the neck, producing non-linear deformations between the images. We modeled the movement of the neck using a "twisting and bending model" with only six parameters for non-rigid registration. We used a Mutual Information (MI) based image similarity measure with the Powell optimization method as they have been used effectively in US image registration applications. For evaluation of our algorithm, we acquired 3D US carotid images from three subjects at two different head positions to simulate images acquired at different times. Then, we registered each image set using our "twisting bending model" based non-rigid registration algorithm. We calculated the Mean Registration Error (MRE) between the segmented vessel surfaces in the target image and the registered image using a distance-based error metric. We repeated the experiment with only rigid registration to compare the capabilities of the proposed algorithm in improving registration of 3D carotid US images. The average registration error was 1.03+/-0.23 mm using our non-rigid registration technique, while it was 1.50+/-0.50 mm when we applied the rigid registration alone.