Prediction of Optimal MRI Protocols Using Seq2seq with Attention and ICD-10 Code

Abstract

　The purpose of this study was to evaluate the accuracy of optimal MRI protocols using a Seq2seq with attention and ICD-10-based classification on the information provided in MRI examination request forms. A total of 6,831 contrast-enhanced brain MRI forms and 10,754 non-contrast-enhanced brain MRI forms were collected. Three methods were used to predict MRI protocols: (1) Seq2seq with attention, (2) ICD-10 code classification, and (3) the combination of these two methods. For the Seq2seq with attention, tokenization was performed using three methods: unigram, bigram, and morpheme. It was observed that, for both contrast-enhanced and non-contrast examinations, the highest complete coverage rate of the radiologist’s instructions was achieved using method (3) with bi-gram (60.35% and 66.16%, for contrast-enhanced and non-contrast-enhanced MRI, respectively). The respective rates were 51.67% and 54.66% for method (1) using bi-gram, and 1.64% and 39.46% for method (2). Because the complete coverage rate increased by 10-20% compared to the individual methods, it is thought that the combination of Seq2seq and ICD-10 codes was the most useful method for predicting optimal MRI protocols.