Automatic Detection of Thyroid and Adrenal Incidentals Using Radiology Reports and Deep Learning

Abstract

Background: Computed tomography (CT) is commonly performed when evaluating trauma patients with up to 55% showing incidental findings. Current workflows to identify and inform patients are time-consuming and prone to error. Our objective was to automatically identify thyroid and adrenal lesions in radiology reports using deep learning.

Materials and methods: All trauma patients who presented to an accredited Level 1 Trauma Center between January 2008 and January 2019 were included. Radiology reports of CT scans that included either a thyroid or adrenal gland were obtained. Preprocessing included word tokenization, removal of stop words, removal of punctuation, and replacement of misspellings. A word2vec model was trained using 1.4 million radiology reports. Both training and testing reports were selected at random, manually reviewed, and were considered the gold standard. True positive cases were defined as any lesions in the thyroid or adrenal gland, respectively. Training data was used to create models that would identify reports that contained either thyroid or adrenal lesions. Our primary outcomes were sensitivity and specificity of the models using predetermined thresholds on a separate testing dataset.

Results: A total of 51,771 reports were identified on 35,859 trauma patients. A total of 1,789 reports were annotated for training and 500 for testing. The thyroid model predictions resulted in a 90.0% sensitivity and 95.3% specificity. The adrenal model predictions resulted in a 92.3% sensitivity and a 91.1% specificity. A total of 240 reports were confirmed to have thyroid incidentals (mean age 69.1 yrs ± 18.9, 35% M) and 214 reports with adrenal incidentals (mean age 68.7 yrs ± 16.9, 50.5% M).

Conclusions: Both the thyroid and adrenal models have excellent performance with sensitivities and specificities in the 90s. Our deep learning model has the potential to reduce administrative costs and improve the process of informing patients.

Keywords: Deep learning; Incidental findings; Natural language processing; Trauma centers.