A Novel Method for Audiogram Digitization in Audiological Reports

Abstract

An audiogram records the hearing status, including each hearing threshold at multiple frequencies. While deep learning is gradually maturing in the clinical research approach, audiologists could speed up their diagnostic process with audiogram digitization for handwritten graphs or electronically generated images from instruments. However, given the diversity of audiogram symbols and formats, the existing audiogram digitization model has room for improvement in recognition accuracy. We propose a multi-stage workflow to enhance accuracy by integrating YOLOv5 and the optical character recognition (OCR) model. Our proposed audiogram digitization model could identify all audiogram symbols with an accuracy rate of 98%. We hope that this model could help future research in audiology.