Enhancing Radiology Documentation using a Lightweight CNN-RNN based Speech-to-Text Model

The healthcare system is witnessing a tremendous rise in medical data due to technological adoption in the medical field. Documenting the health records, such as radiology reports, lab reports, and clinical interactions, is imperative for better healthcare delivery. During the COVID pandemic, radiologists attended to a large number of patients in a day, and it turned out to be a burnout for them. This study aims to automate the radiology report generation process to reduce documentation time. Documenting the fine details during examination by radiologists alleviates the physician’s doubts. Traditionally, radiologists either dictate their examination to a medical scribe, or these observations are documented after the examination. This study focuses on potential benefits of using Speech-to-Text (STT) technology to streamline radiology workflows, reduce clinical report generation time, and enhance clinical care. For this study, a lightweight custom deep learning model, Convolutional Neural Network with Recurrent Neural Network (CNN-RNN), was developed to transcribe radiology audio sentences to text. A dataset of 100 radiology sentences from Kaggle dataset, which is a representative of real-world radiology sentences, was used for the evaluation. The accuracy of the model was assessed using standard metrics including word error rate (WER), Medical Term Identification Error (MTIE), and overall transcription accuracy. The system achieved $94 \%$ transcription accuracy from audio sentences with a limited dataset. Also, the system performed better in identifying medical terms with almost 99% accuracy among the 48 medical terms from 100 sentences. From the experimental study, it is evident that the usage of STT is a promising technology for the radiologists to prepare radiology reports without any delay.