>

072. A framework for using machine learning algorithms to detect ventriculoperitoneal shunt valve settings from skull radiographs

Authors: Swetha Sundar, MD

Introduction:
Identification of shunt valve types and settings is typically performed visually by neurosurgeons using skull radiographs. This process is time intensive and subject to human error. We are developing machine learning algorithms to automate this process. We created a multistep framework consisting of 1)image processing and feature generation, 2)identification of valve make/model, 3)valve segmentation and isolation, and 4)detection of valve setting. We present our analysis of valve isolation and segmentation through a deep learning geometric approach.Methods: In this IRB-approved project, we used adult and pediatric skull radiographs and manually identified and cropped radiographs of 10 valves of various settings of a single brand. From each image, we generated 7 additional settings using a rotational matrix calculated from the brand’s documentation of all possible settings and a mask of the valve’s inner core. Eighty total cropped images were generated. Data were de-noised using specific thresholding of background density, then normalized, and reduced to 64x64 pixel images. Masks were generated for each image in a triangular geometry (two inner valve components and reference point). Of the data, 80% was used for the training group and 20% for the testing group. TensorFlow was used for the deep learning network and U-Net was used to perform the segmentation.
Results:
The segmentation model was trained for 60 epochs. Accuracy in identifying the correct mask was 99% in the training group and 94% in the testing group. This was confirmed by visual inspection of the mask. The image would then be carried on for triangulation of the mask to represent the underlying shape of the valve setting.
Conclusion:
We were able to isolate and segment a single brand of shunt valve using a deep learning geometric approach. This is a critical component of a multistep framework that will be used to fully automate shunt detection in the future.