Autonomous diagnosis of pediatric cutaneous vascular anomalies using a convolutional neural network

Palak Patel, Louisiana State University Health Sciences Center, Department of Otolaryngology, Head and Neck Surgery, New Orleans, Louisiana, USA. Electronic address: ppat13@lsuhsc.edu.
Katelyn Ragland, University of Arkansas for Medical Sciences, Department of Otolaryngology, Head and Neck Surgery, Little Rock, Arkansas, USA. Electronic address: kmragland@uams.edu.
Brianna Robertson, Louisiana State University School of Engineering, Baton Rouge, Louisiana, USA. Electronic address: brobe86@lsu.edu.
Gabriel Ragusa, Louisiana State University School of Engineering, Baton Rouge, Louisiana, USA. Electronic address: gragus2@lsu.edu.
Christine Wiley, Louisiana State University School of Engineering, Baton Rouge, Louisiana, USA. Electronic address: cwile12@lsu.edu.
Jacob Miller, Louisiana State University School of Engineering, Baton Rouge, Louisiana, USA. Electronic address: jmil266@lsu.edu.
Robert Jullens, Louisiana State University School of Engineering, Baton Rouge, Louisiana, USA. Electronic address: rjulle3@lsu.edu.
Michael Dunham, Louisiana State University Health Sciences Center, Department of Otolaryngology, Head and Neck Surgery, New Orleans, Louisiana, USA. Electronic address: mdunha@lsuhsc.edu.
Gresham Richter, University of Arkansas for Medical Sciences, Department of Otolaryngology, Head and Neck Surgery, Little Rock, Arkansas, USA. Electronic address: gtrichter@uams.edu.

Abstract

OBJECTIVES: Design and validate a novel handheld device for the autonomous diagnosis of pediatric vascular anomalies using a convolutional neural network (CNN). STUDY DESIGN: Retrospective, cross-sectional study of medical images. Computer aided design and 3D printed manufacturing. METHODS: We obtained a series of head and neck vascular anomaly images in pediatric patients from the database maintained in a large multidisciplinary vascular anomalies clinic. The database was supplemented with additional images from the internet. Four diagnostic classes were recognized in the dataset - infantile hemangioma, capillary malformation, venous malformation, and arterio-venous malformation. Our group designed and implemented a convolutional neural network to recognize the four classes of vascular anomalies as well as a fifth class consisting of none of the vascular anomalies. The system was based on the Inception-Resnet neural network using transfer learning. For deployment, we designed and built a compact, handheld device including a central processing unit, display subsystems, and control electronics. The device focuses upon and autonomously classifies pediatric vascular lesions. RESULTS: The multiclass system distinguished the diagnostic categories with an overall accuracy of 84%. The inclusion of lesion metadata improved overall accuracy to 94%. Sensitivity ranged from 88% (venous malformation) to 100% (arterio-venous malformation and capillary malformation). CONCLUSIONS: An easily deployed handheld device to autonomously diagnose pediatric skin lesions is feasible. Large training datasets and novel neural network architectures will be required for successful implementation.