Contact Us
CiiS Lab
Johns Hopkins University
112 Hackerman Hall
3400 N. Charles Street
Baltimore, MD 21218
Directions
Lab Director
Russell Taylor
127 Hackerman Hall
rht@jhu.edu
Last updated: 4/20/2020 at 11:56PM
This project concerns developing and validating an image guidance framework for application to a robotic-assisted fibular reduction in ankle fracture surgery. The aim is to produce and demonstrate proper functioning of software for automatic determination of directions for fibular repositioning with the ultimate goal of application to a robotic reduction procedure that can reduce the time and complexity of the procedure as well as provide the benefits of reduced error in ideal final fibular position, improved syndesmosis restoration and reduced incidence of post-traumatic osteoarthritis. The focus of this product will be developing and testing the image guidance software, from the input of preoperative images through the steps of automated segmentation and registration until the output of a final transformation that can be used as instructions to a robot on how to reposition the fibula, but will not involve developing or implementing the hardware of the robot itself.
Ankle fractures occur with a frequency of around 174 cases per 100,000 adults per year, with over 5 million yearly cases in the U.S. alone (Goost et al), affecting mainly young active people and the elderly. Ankle fractures most commonly involve a fracture in the lower fibula which can also result in disruption of the syndesmosis, or the alignment of other bones and ligaments within the ankle joint, if the fibula is displaced. This is due to the displacement of the lower fibula causing damage and forceful shifting of these ligaments and other connective tissue. This can lead to many long-term complications including post-traumatic osteoarthritis (PTOA) if the proper syndesmosis is not restored.
The proximal fracture can be treated with relatively little difficulty such as by fixation with metal screws. However, the second part of the surgery is fibular reduction, or movement of the lower fibula back into proper position to reconstruct the joint and restore syndesmosis. The current standard of care for this procedure is visual estimation by the surgeon to determine where to place the fibula and screw-based fixation of syndesmosis. In one study over 20% of patients who underwent reduction via this method were shown through CT scans shortly afterwards to have syndesmotic malreduction, which was defined as a greater than 2 mm widening of syndesmosis compared to the patient’s other, healthy ankle (Nagvi et al). The most common reason for re-operation in the weeks immediately following the initial ankle fracture surgery is syndesmotic malreduction (Ovaska et al). Furthermore, the incidence of PTOA in ankle fracture patients is as high as 70% (Mehta et al). Evidently, the standard of care for reduction is inadequate for proper syndesmotic restoration and prevention of PTOA.
Steps in Approach:
1. Automated Segmentation
The first step in an image-guided approach is fast, automated segmentation of the pre-operative images in order to identify and save the relevant anatomical features, in this case the fibula and other bones of the ankle. A promising existing approach to automated segmentation is active shape models (ASMs) (Brehler et al). An ASM model is trained by taking in as input a training set of several segmentations of a single bone from different patients and then performing a PCA-based analysis to identify and characterize the principal modes of variation in the surface morphology of the bone within the population of training samples. The model comprises a set of basis functions for each principal component of variation, which can be tuned and create a deformable template that can be mapped onto any new example by minimizing residuals. Thus, this can be applied to automated segmentation by using this template object in order to map onto a new image and segment the bone given some initialization.
ASMs can be improved even further by updating them to Coupled Active Shape Models (cASMs). Traditional ASMs are trained independently for individual bones and separately initialized, however as a result they are susceptible to errors in the narrow articular joint spaces between bones if there is poor contrast in the image and can result in producing segmentations with overlap in adjacent bones. cASMS are trained with consideration of multiple bones simultaneously and incorporate the spatial relations and articular joint widths between adjacent bones as features as well, and thus they are able to improve on the accuracy of segmentation within these joint spaces and prevent overlaps.
While in this work the cASMs were evaluated in high-resolution cone-beam CT (CBCT) images, they have not yet been tested in pre-operative C-arm images for 2D-3D reconstruction. Thus, an important step in adopting the cASM approach for this project will be to evaluate its accuracy in segmenting C-arm images and improving it to do so, including improving the initializations and adding corrections for noise and artifacts.
A deep learning approach will also be taken, as it is likely that the ASM model will not provide sufficient segmentation accuracy. It will be used to also conduct image segmentation, like the ASM. The purpose of this is to build this model to hopefully get a higher segmentation accuracy than the ASM model. The final approach chosen will be the one that yields the highest accuracy. It is possible that the final model chosen is a combination of the ASM and the neural network. The plan for the deep neural network is to utilize a convolutional neural network (CNN). These are commonly used for image processing in a lot of domains. The specifics of the neural network are currently unknown. However, there are some major characteristics that will most likely be implemented. Maximum pooling will most likely be used over average pooling as it is known to be more effective in image classification. This is because this technique focuses on the most important characteristics in the image. This is helpful in classification because it does not let other, less significant, pixels affect the classification and only looks at what is really important. Pooling is also beneficial because it is able to focus on what is really important. This reduces the computational burden of training the neural net. Other techniques such as dropout. Another strategy that could be used is increasing the number of epochs. This will need to be carefully managed because if intraoperative c-arm data is being utilized, the segmentation will need to be determined quickly. The model will need to be fine-tuned, which will be an ongoing process. The goal is to achieve high accuracy, the exact accuracy goals for the deep learning model can be seen in the deliverables section.
Whichever final approach is used for automated segmentation, it would be used to automatically segment the ankle bones of both of the patient’s ankles (the healthy one and the injured one). The model, or combination of models, which yields the highest segmentation accuracy (as compared to manually segmented images through minimizing square residuals) will be used.
2. Metal Artifacts
After segmenting healthy ankles, ankles with metal artifacts will be looked at. Since there is a lack of data with metal artifacts and it is not currently possible to obtain more of it, metal reduction algorithms will be applied to this data. Then this data will be incorporated into the final segmentation model.
[1] M. Brehler, A. Islam, L. Vogelsang, D. Yang, W. Sehnert, D. Shakoor, S. Demehri, J. H. Siewerdsen, W. Zbijewski, “Coupled Active Shape Models for Automated Segmentation and Landmark Localization in High-Resolution CT of the Foot and Ankle”
[2] Goost H, Wimmer MD, Barg A, Kabir K, Valderrabano V, Burger C. Fractures of the ankle joint: investigation and treatment options. Dtsch Arztebl Int. 2014;111(21):377–388. doi:10.3238/arztebl.2014.0377
[3] Ovaska MT, Mäkinen TJ, Madanat R, Kiljunen V, Lindahl J. A comprehensive analysis of patients with malreduced ankle fractures undergoing re-operation. Int Orthop. 2014;38(1):83–88. doi:10.1007/s00264-013-2168-y
[4] Mehta, Saurabh Sagar, et al. “Understanding Risks and Complications in the Management of Ankle Fractures.” Indian Journal of Orthopaedics, Medknow Publications & Media Pvt Ltd, Sept. 2014, www.ncbi.nlm.nih.gov/pmc/articles/PMC4175856/.
[5] Naqvi, G. A., Cunningham, P., Lynch, B., Galvin, R., & Awan, N. (2012). Fixation of Ankle Syndesmotic Injuries: Comparison of TightRope Fixation and Syndesmotic Screw Fixation for Accuracy of Syndesmotic Reduction. The American Journal of Sports Medicine, 40(12), 2828–2835. https://doi.org/10.1177/0363546512461480
[6] Y. Otake, R. J. Murphy, R. B. Grupp, Y. Sato, R. H. Taylor, and M. Armand, “Comparison of Optimization Strategy and Similarity Metric in Atlas-to-subject Registration Using Statistical Deformation Model”, in Proc. SPIE 9415, Medical Imaging 2015: Image-Guided Procedures, Robotic Interventions, and Modeling, San Francisco, 8-10 Feb., 2015. p. 94150Q. 10.1117/12.2081310
[7] Weng, Y., Zhou, T., Li, Y., & Qiu, X. (2019). NAS-Unet: Neural Architecture Search for Medical Image Segmentation. IEEE Access, 7, 44247–44257. doi: 10.1109/access.2019.2908991
[8] Kuo-Sheng Cheng, Jzau-Sheng Lin and Chi-Wu Mao, “The application of competitive Hopfield neural network to medical image segmentation,” in IEEE Transactions on Medical Imaging, vol. 15, no. 4, pp. 560-567, Aug. 1996.
[9] Xu, S., Uneri, A., Jay Khanna, A., et al. Polyenergetic known-component CT reconstruction with unknown material compositions and unknown x-ray spectra. Physics in Medicine & Biology 62 (8), 3352–3374 (2017).
