MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

A Constrained Adversarial Deep Neural Network Based Approach for Highly Accurate and Real Time Patient Safety Workflow

A Santhanam*, Y Min , N Agazaryan , P Beron , D Low , UCLA, Los Angeles, CA

Presentations

(Monday, 7/30/2018) 1:15 PM - 1:45 PM

Room: Exhibit Hall | Forum 5

Purpose: Catastrophic errors such as wrong patient positioning and wrong patients getting treated still occur in radiotherapy, in spite of employing state-of-art manual and semiautomatic verification methods. An automated framework for detecting such errors is needed.

Methods: In this abstract, we present a constrained adversarial deep neural network for automatically identifying objects of interest (e.g. patient body) in the treatment space, for developing clinical workflows focused on ensuring patient safety. Inputs to this neural network stem from a multi-3D camera based setup was first assembled and instantiated within a radiotherapy setup for imaging the treatment space. The 3D regions representing the patient as well as the treatment equipment needed to be classified for learning purposes. Generating precise training data was critical to ensure a successful recognition of the patient’s 3D contour. To this end, we employed a GPU-based grab-cut implementation to generate the patient body contour. The training process took the 3D treatment space view from every 3D camera and the segmented entity as data and labels, respectively. The adversarial network generated the label (object of interest) for the given data (3D treatment space view). The training process was completed using the data/label generated for a 4-day time period (total number of 3D snapshots was 4,800). Once trained, the network automatically generated the patient 3D contour and the treatment space contour from each 3D treatment space view.

Results: The network was tested for data consisting for 40 days (48,000 3D snapshots). As the images were obtained in a conventional clinical setup, the data included variations in the patient posture and setup, variations in the gantry positions, and presence of therapists in the treatment space with 98% accuracy.

Conclusion: An automated patient safety workflow can be achieved in real-time using a set of 3D cameras and a machine learning approach.

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email