US 11,915,809 B2
Method and system for automatically generating a section in a radiology report
Jeffrey Chang, Berkeley, CA (US); Doktor Gurson, Berkeley, CA (US); Eric Purdy, Berkeley, CA (US); Brandon Duderstadt, Berkeley, CA (US); Jeffrey Snell, Berkeley, CA (US); Andriy Mulyar, Berkeley, CA (US); and Deeptanshu Jha, Berkeley, CA (US)
Assigned to RAD AI, Inc., Berkeley, CA (US)
Filed by RAD AI, Inc., Berkeley, CA (US)
Filed on Jun. 9, 2023, as Appl. No. 18/207,880.
Application 18/207,880 is a continuation of application No. 17/725,031, filed on Apr. 20, 2022.
Application 17/725,031 is a continuation of application No. 17/020,593, filed on Sep. 14, 2020, granted, now 11,342,055, issued on May 24, 2022.
Claims priority of provisional application 62/900,148, filed on Sep. 13, 2019.
Prior Publication US 2023/0343426 A1, Oct. 26, 2023
This patent is subject to a terminal disclaimer.
Int. Cl. G16H 15/00 (2018.01); G16H 10/60 (2018.01); G16H 20/40 (2018.01); G16H 40/63 (2018.01); G16H 40/67 (2018.01); G16H 50/20 (2018.01)
CPC G16H 15/00 (2018.01) [G16H 10/60 (2018.01); G16H 20/40 (2018.01); G16H 40/63 (2018.01); G16H 40/67 (2018.01); G16H 50/20 (2018.01)] 16 Claims
OG exemplary drawing
 
1. A method, comprising:
training a machine learning model based on a first set of historical radiology reports, wherein the machine learning model comprises a transformer model comprising a set of encoders and a set of decoders;
tuning the machine learning model based on a second set of historical radiology reports, wherein the second set of historical radiology reports are associated with a radiologist, wherein tuning the machine learning model comprises learning a writing style for the radiologist, the writing style reflecting at least one of: a length metric, a word choice, an ordering style, or a summarization style;
determining a set of finding inputs for a radiology report associated with a patient;
using the set of encoders of the tuned model, determining a set of embeddings based on the set of finding inputs, and determining a context matrix based on the set of embeddings;
using the set of decoders of the tuned model, generating an impression section of the radiology report based on the radiologist style matrix and the context matrix, wherein the impression section is configured to mimic the writing style for the radiologist;
retraining the machine learning model based on the generated impression section of the radiology report to determine a retrained model; and
with the retrained model, generating a second impression section of a second radiology report.