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: 05/05/2020 12:00
Designing a deep learning-based approach improving K-wire identification in fluoroscopic images of the pelvis with the goal of reducing patient radiation exposure and improved accuracy of K-wire insertion in pelvic stabilization surgeries.
Our proposed approach is outlined below:
Pelvic fractures occur across all demographics and are estimated to be 2-8% of all fractures in the UnitedStates [1]. They are, however, primarily occurring in younger demographics due to high energy trauma such as traffic accidents and in older demographics due to day-to-day falls. Osteoporosis often causes bones to get weaker with age, and thus pelvic fractures are becoming more common as the population of theUnited States is aging [2]. As a result, pelvic fracture treatment is a growing market and is estimated to reach $1.8 billion by 2025 [3]. Although some cases can be treated with external stabilization, unstable pelvic fractures require surgical stabilization. Minimally invasive surgical (MIS) procedures are becoming an increasingly popular choice for some surgeons as they decrease blood loss, risk of infection, and recovery times for patients with unstable pelvic fractures compared to the open surgery approaches [4]. During MIS procedures, Kirschner wires (K-wires), which are thin, foot-long stainless steel pointed rods, are used to guide cannulated screws into place to join pelvic bone fragments. Inserting K-wires is challenging even for experienced surgeons since pelvis has a complex structure and K-wires are flexible and thus bend inside the patient. Surgeons, therefore, cannot use conventional marker-based navigation methods that assume a rigid guided object and are not able to accurately guide the trajectory of the K-wire. To guide the K-wire, surgeons rely on intraoperative fluoroscopy, taking images throughout the insertion of the K-wire to visualize its position relative to the patient’s anatomy. As a consequence, patients often get exposed up to 2 minutes of radiation per screw [5]. Despite the extensive imaging, however, the surgeons still struggle to achieve the permissible accuracy for the K-wire’s trajectory of 1 mm translational error and5◦rotational error [6]. Since the K-wire is used to place stabilization screws, screw malposition is a common complication of pelvic fracture stabilization surgeries: 20-30% screw placements are rated as suboptimal [7] and 6% breach the cortical bone [8]. Screw malposition is a serious issue as it can result in neurological and vascular injuries, require longer surgical times, and lead to long-term pelvic instability [9].
There are K-wire navigation systems getting developed, but even novel solutions to K-wire detection and identification are slow and lack the accuracy. Therefore, our project hopes to improve K-wire insertion accuracy to reduce injury while limiting radiation exposure via the following aim.
In the existing K-Wire placement process, patients are often exposed to high dosages of radiation for extended periods of time due to difficulties in localization. The development of a prediction model to improve the detection speed while maintaining high detection accuracy could reduce both exposure to radiation and error due to visibility limitations.
Access to quality, annotated training datasets remains a significant pain point for the application of deep learning to medical image analysis. The benefits of bringing such technology to medicine is high, but it is dependent on the availability of large amounts of annotated data. Therefore we propose a transfer learning approach based on fully simulated training data.
Our training dataset will consist of real fluoroscopic pelvic images augmented with simulated K-wires. It will include augmenting fluoroscopic images of the pelvis with both rigid and deformed K-wires. The K-wire trajectories will be randomly selected. By simulating the K-Wire, we are able to easily obtain the segmented mask as well. We observed a trend that with increased background variation (increasing N in the figure below), we are able to obtain better performance on the test set. Thus in future work, we hope to perform an in-depth analysis into the number of input background images versus performance.
For this initial analysis, we limit ourselves to AP-like views of the pelvis with single K-Wires in order to constrain the scope of the problem. Given our success, we now hope to test whether our data generation and model will also have similar performance on a broader set of views.
The end to end data generation process can be seen below:
Implemented U-Net [18] and U-Net++ [19] networks evaluated performance on K-Wire detection task. The model architectures were obtained from code open sourced by [19] and adapted to fit our task. We evaluated impact to performance by adapting the following parameters:
In future work, we are interesting in evaluating a more lightweight model, such as U-Net Light [20] as well as compare U-Net based architectures to Mask R-CNN architectures.
To evaluate the success of our transfer learning approach, we tested our trained model on both a simulated dataset as well as on a real test dataset of 33 images with K-wires. We leveraged 4 standard CNN metrics (area under curve, precision, recall, F1 Score) and a task specific centerline detection metric using Hausdorff distance to quantify performance.
Standard CNN Prediction Metrics Calculation:
K-Wire Detection Task Specific Metrics Calculation Process:
This maximum deliverable has been pushed to the summer. Please see our project proposal to view initial thoughts.
KWire Detection GitLab Repository
**README.md files within the code repository provide further explanations
Here give list of other project files (e.g., source code) associated with the project. If these are online give a link to an appropriate external repository or to uploaded media files under this name space (456-2020-09).
