Computer Vision for Cerebrospinal Fluid Valve Detection
Matthew Grabowski, Pablo Recinos, Ghaith Habboub, David Piraino, Violette Recinos, Shahed Tish, Stephen Jones
Introduction: Hydrocephalus surgical management is one of the most common procedures performed by neurosurgeons. Recognizing the type and setting of cerebrospinal fluid valves can be a demanding task due to the increasing number of patients and the variety of valves. Radiology has been an exciting field for utilizing computer vision.
Objective: The goal of this project is to create a computer vision classifier built into a smart phone app and/or embedded into the medical imaging reading system. The algorithm would detect the type and setting of a cerebrospinal fluid valve and auto-populate it into the radiology report.
Methods: We have images on ~ 6500 shunt series. Of these we used only 1.5% of the data and we are in the process of obtaining the rest for analysis. We created a multi-layer convolutional neural network utilizing Tensorflow/Keras, written in Python. We converted the model into an iOS app using CoreML, written in Swift.
Results: A multi-layer deep convolutional neural network was designed. Model accuracy was 60% despite utilizing 1.5% of the data. The iOS application can detect images in real time and provide probabilities utilizing the camera embedded in the smart phone. We were able to achieve acceptable accuracy which we expect to significantly improve it after training the model with the remaining data.
Conclusions: This is the first cerebrospinal fluid valve detection classifier. Valve detection can be achieved with reasonable accuracy using computer vision. This would play an integral role in the future of radiology and decrease the work load on both the surgeon and the radiologist