A machine learning framework for context specific collimation and workflow phase detection

Collimators control the field of view by using thick blades to block x-ray leaving the source to image the patient. When the blades are adjusted, the area of the patient receiving radiation is reduced or increased. Current fluoroscopy systems allow only for manual collimation by the operator. This can be done from the control panel using buttons or joystick. Nonetheless, manual collimation is time consuming, causes interruption to the clinical work-flow, and is operator dependant. This is because the operator has to first identify a region of interest (RoI), then collimate around the RoI depending on the type of the procedure, work-flow phase, and interventionist‘s preferences. In this work, we propose a learning based framework that can autonomously predict the work-flow phase and localize an object of interest during congenital cardiac interventions (CCI) procedures. In particular, we propose to learn the task of work-flow recognition by using a convolutional neural network model. For training and evaluating our model, 4554 images from 25 clinical cases acquired during Biplane CCI procedures at Evelina
London Children’s Hospital, UK were used. The framework allows for optimal collimation to be automatically adjusted in varying amounts depending on the predicted work-flow around the localized devices, which we refer to as context specific collimation.