Dataset Information

Multi-task weak supervision enables anatomically-resolved abnormality detection in whole-body FDG-PET/CT.

ABSTRACT: Computational decision support systems could provide clinical value in whole-body FDG-PET/CT workflows. However, limited availability of labeled data combined with the large size of PET/CT imaging exams make it challenging to apply existing supervised machine learning systems. Leveraging recent advancements in natural language processing, we describe a weak supervision framework that extracts imperfect, yet highly granular, regional abnormality labels from free-text radiology reports. Our framework automatically labels each region in a custom ontology of anatomical regions, providing a structured profile of the pathologies in each imaging exam. Using these generated labels, we then train an attention-based, multi-task CNN architecture to detect and estimate the location of abnormalities in whole-body scans. We demonstrate empirically that our multi-task representation is critical for strong performance on rare abnormalities with limited training data. The representation also contributes to more accurate mortality prediction from imaging data, suggesting the potential utility of our framework beyond abnormality detection and location estimation.

SUBMITTER: Eyuboglu S

PROVIDER: S-EPMC7994797 | biostudies-literature | 2021 Mar

REPOSITORIES: biostudies-literature

ACCESS DATA

Publications

Multi-task weak supervision enables anatomically-resolved abnormality detection in whole-body FDG-PET/CT.

Eyuboglu Sabri S Angus Geoffrey G Patel Bhavik N BN Pareek Anuj A Davidzon Guido G Long Jin J Dunnmon Jared J Lungren Matthew P MP

Nature communications 20210325 1

Computational decision support systems could provide clinical value in whole-body FDG-PET/CT workflows. However, limited availability of labeled data combined with the large size of PET/CT imaging exams make it challenging to apply existing supervised machine learning systems. Leveraging recent advancements in natural language processing, we describe a weak supervision framework that extracts imperfect, yet highly granular, regional abnormality labels from free-text radiology reports. Our framew ...[more]

PMID: 33767174

Similar Datasets

Project description:BACKGROUND:Hyperglycaemia can influence 18F-fluorodeoxyglucose (FDG) uptake due to competition for glucose transport and phosphorylation by hexokinase. Major international nuclear medicine societies recommend blood glucose level (BGL) < 11.1 mmol/L (200 mg/dL) prior to performing FDG positron emission tomography/computed tomography (PET/CT). However, there is no consensus approach and complications of previously proposed insulin guidelines included significant hypoglycaemia, inconvenience and skeletal muscle uptake. This study aims to establish the safety and efficacy of a personalised insulin calculator protocol to estimate the dose of intravenous insulin injection for correction of hyperglycaemia prior to FDG PET/CT. RESULTS:This is a retrospective audit of all patients treated with insulin for hyperglycaemia (BGL > 10 mmol/L) prior to FDG PET/CT at the Peter MacCallum Cancer Centre over a 2-year period. Cohort 1 comprised a 12-month period (April 1, 2014-March 31, 2015) using the department's established empiric-dose insulin protocol, and Cohort 2 the 12 months (April 1, 2015-March 31, 2016) following introduction of a personalised insulin calculator protocol. Variables including body mass index, insulin-dose calculated and/or administered, BGL at baseline and nadir, and time to FDG injection were analysed. There were 115 and 136 patients treated with insulin in Cohorts 1 and 2 respectively, with similar baseline variables including mean BGL (14.5 vs 14.4 mmol/L) and range (10.5-22.7 vs 10.4-24.3 mmol/L). Use of the new personalised insulin calculator resulted in significantly fewer hypoglycaemic events (0.7% vs 5.2%; P < 0.03), shorter median time from insulin to FDG injections (108 min vs 136 min; P < 0.001) and greater individualised range in insulin prescription (3-32 IU vs 4-20 IU). The majority of patients (88.3%) receiving the personalised insulin calculator prescribed dose achieved BGL < 10.0 mmol/L. All scans obtained were of diagnostic quality. CONCLUSIONS:The use of our personalised insulin calculator protocol effectively lowered BGL to the target range, resulted in significantly fewer hypoglycaemic events and reduced median time between insulin and FDG injection compared to a pre-existing empiric protocol whilst achieving diagnostic scans.

Dataset Information

Multi-task weak supervision enables anatomically-resolved abnormality detection in whole-body FDG-PET/CT.

Publications

Multi-task weak supervision enables anatomically-resolved abnormality detection in whole-body FDG-PET/CT.

Similar Datasets

OmicsDI is part of the ELIXIR infrastructure

Tweets