PSRC Main Site  |  Past & Future Meetings
Plastic Surgery Research Council

Back to 2022 Abstracts


Classification Of Plagiocephaly Severity Using Machine Learning Algorithms
Huan Nguyen, B.S., Ellen Wang, B.S., Phuong Nguyen, M.D., Matthew Greives, M.D..
McGovern Medical School at UTHSC Houston, Houston, TX, USA.

PURPOSE: With a recent increase in incidence of pediatric plagiocephaly, there is a need for a reliable and accurate method to diagnose cranial abnormalities and monitor treatment progress in patients. The current standard is the Argenta scale classification (AS), which comprises a scale of 1-5 with clinical diagnostic criteria including posterior, frontal, and facial asymmetry, ear malposition, and posterior vertical cranial growth.1, 3 A machine learning (ML) algorithm would provide physicians with objective classification and gather more comprehensive data than subjective observation. The purpose of our study is to build an artificial intelligence (AI) program that provides physicians and patient families reliable and dynamic data to formulate an initial diagnosis, and ultimately, predict treatment time and track treatment progress.
METHODS: 2D photographs were taken of the vertex and AP view on infants who were assessed for plagiocephaly. A custom-built AI program was created to automatically identify the head circumference from the image and calculate cranial vault asymmetry index (CVAI) at multiple angles from the anterior-posterior line.2 Using these measurements, the program applied an ML training model built from images of previous patients with known AS severities classified by an experienced craniofacial surgeon to classify an AS for the given image. The resulting AS predicted by the program was compared to an experienced craniofacial surgeonís diagnosis to assess for accuracy.
RESULTS: 19 patients were assessed from a single institution. The program used ML methods to create thresholds for different classes of the AS. 3 out of 3 (100%) with AS 1, 4 out of 6 (67%) with AS 2, 4 out of 6 (67%) with AS 3, and 3 out of 4 (75%) with AS 4 were correctly identified by the AI program. AS 5 could not be determined due to the limitation of the program analyzing only vertex view images.
CONCLUSION: Using machine learning algorithms, we were able to create objective measurements to classify the Argenta scale. This will enable the clinician to make more data driven recommendations for therapy and formulate a customized prognosis and treatment plan such as when to use a molding helmet.


Back to 2022 Abstracts