Abstract Archives of the RSNA, 2014
Gady agam PhD, Presenter: Shareholder, RaPID Medical Technologies, LLC
Vicko Vicko Gluncic, Abstract Co-Author: Founder, RaPID Medical Technologies, LLC
CEO, RaPID Medical Technologies, LLC
Serge Kobsa MD, PhD, Abstract Co-Author: Nothing to Disclose
Shirley Richard MBA, Abstract Co-Author: Clinical Advisor, RaPID Medical Technologies, LLC
Mario Moric, Abstract Co-Author: Officer, RaPID Medical Technologies
Sameer A. Ansari MD, PhD, Abstract Co-Author: Shareholder, RaPID Medical Technologies, LLC
RSI are commonly surgical sponges/needles inadvertently left in a patient's body (0.02-1% incidence). They are a priority OR patient safety concern, classified as sentinel events by the Joint Commission. Measures to prevent RSI include effective OR communication, mandatory counts of surgical instruments/sponges, methodical wound examination, and XR. Miscounts occur in <12.5% of surgeries, requiring XR of the surgical field. Since 80-90% of RSI occur with “correct surgical counts” hospitals may mandate XR at the end of all complex surgeries. Although XR based protocols are crucial for RSI detection, they are limited by the sensitivity and experience of the human eye, lack of formal training for RSI detection, and a time intensive process for complete analysis. These limitations motivated us to develop CADe software for the ray-tec XR detectable sponge.
Developed CADe software involves three steps: enhancement, detection, and recognition. Enhancement reduces noise and improves contrast to assist detection. Identification detects candidate regions and classifies them using a supervised learning technique based on extracted 112 dimensional feature vectors. Detected regions are clustered to form candidates using a spatial clustering algorithm. The object detection stage uses an additional classifier that is trained in a supervised manner.
The test collection included a total of 790 images where 277 have one or more ray-tec sponges. We manually labeled the images by marking the area of each sponge and automatically marked negative candidate locations. Overall we had 561 positive locations and 25,638 negative locations.10 fold cross validation was performed and receiver operating characteristic curve generated. Using the optimal point on the curve we obtained 509 true positives, 25,611 true negatives, 27 false positives, and 52 false negatives. CADe beta prototype testing resulted in 99% specificity, 90% sensitivity, and 0.92 F-measure.
Our data demonstrate the feasibility of the CADe for the ray-tec sponge in XRs based on supervised learning methods. Developed CADe software has a potential to improve OR time utilization, RSI detection rate, and overall patient safety.
agam, G,
Vicko Gluncic, V,
Kobsa, S,
Richard, S,
Moric, M,
Ansari, S,
Computer Aided Detection (CADe) of Retained Surgical Items (RSI) in X-ray Images (XR). Radiological Society of North America 2014 Scientific Assembly and Annual Meeting, - ,Chicago IL.
http://archive.rsna.org/2014/14015233.html