Overview

Schema Version

https://atlas.rsna.org/schemas/2025-11/model.json

Name

Deep Learning-based Assessment of Oncologic Outcomes from Structured Radiology Reports

Link

https://dx.doi.org/10.1148/ryai.220055

Indexing

Keywords: Natural language processing, Radiology reports, Oncology, Tumor response category, RECIST, BERT, German language, Free-text reports, Structured reporting

Content: IN, OI, CT, MR, US

RadLex: RID11511, RID10326, RID10312, RID10321, RID11513, RID45910

Author(s)

Matthias A. Fink

Klaus Kades

Arved Bischoff

Martin Moll

Merle Schnell

Maike Küchler

Gregor Köhler

Jan Sellner

Claus Peter Heussel

Hans-Ulrich Kauczor

Heinz-Peter Schlemmer

Klaus Maier-Hein

Tim F. Weber

Jens Kleesiek

Organization(s)

Clinic for Diagnostic and Interventional Radiology, Heidelberg University Hospital

Pattern Analysis and Learning Group, Department of Radiation Oncology, Heidelberg University Hospital

Translational Lung Research Center Heidelberg (TLRC), German Center for Lung Research (DZL)

Faculty of Mathematics and Computer Science, Heidelberg University

Department of Diagnostic and Interventional Radiology with Nuclear Medicine, Heidelberg Thoracic Clinic, Heidelberg University

Division of Medical Image Computing, DKFZ (German Cancer Research Center)

Department of Computer Assisted Medical Interventions (CAMI), DKFZ

Department of Radiology, DKFZ

German Cancer Consortium (DKTK), Partner Sites Essen and Heidelberg

Institute for Artificial Intelligence in Medicine (IKIM), University Medicine Essen

Version

1.0

License

URL: https://pubs.rsna.org/doi/10.1148/ryai.220055

Funding

Authors declared no funding for this work.

Ethical review

IRB-approved retrospective study (Heidelberg University Hospital; approval no. S-083/2018); HIPAA-compliant; informed consent waived.

Date

Updated: 2022-09-01

Published: 2022-07-20

Created: 2022-03-15

References

[1] Fink MA, Kades K, Bischoff A, et al.. "Deep Learning–based Assessment of Oncologic Outcomes from Natural Language Processing of Structured Radiology Reports". Radiology: Artificial Intelligence. 2022;4(5):e220055.. 2022-09-01. doi:10.1148/ryai.220055. PMID: 36204531.

Model

Architecture

Transformer-based language model (BERT) pretrained for German; fine-tuned for multi-class classification of tumor response categories. Reference models: Linear Support Vector Classifier, k-Nearest Neighbors, Multinomial Naive Bayes with TF-IDF features.

Clinical benefit

Automated extraction of oncologic outcomes (RECIST-related tumor response categories) from free-text radiology reports to support longitudinal assessment and tumor board decision-making.

Clinical workflow phase

Clinical decision support systems; workflow optimization for tumor board preparation and outcomes curation.

Degree of automation

Fully automated NLP classification of TRC from report text; no manual feature engineering required.

Indications for use

Classification of tumor response categories (progressive disease, stable disease, partial response, complete response) from German-language radiology free-text oncology reports in patients undergoing oncologic imaging.

Input

German-language radiology reports: training on mined structured oncology reports (SOR) and inference on free-text oncology reports (FTOR).

Limitations

Models trained on German SOR; generalizability to other languages not established. No FTOR used for training in this study; performance affected by lexical complexity and semantic ambiguities. Lower performance for Stable Disease classification and when oncologic and nononcologic findings are discordant. Ground truth labels mined from SOR using regular expressions; 24.8% of SOR excluded due to failed extraction (equivocal cases). Standard of reference based on reports only; no image review or quantitative RECIST tables in FTOR.

Output

Description: Multi-class classification of tumor response category per report; probabilities per class; can be aggregated into longitudinal timelines per patient.

Recommendation

Potential use for automated curation of oncologic outcomes from large volumes of reports and as decision support for multidisciplinary tumor boards.

Reproducibility

Fivefold cross-validation on SOR training set; held-out SOR test subset; detailed hyperparameters provided in supplemental material (Table E1, Figure E1).

Use

Intended: Report data extraction

Out-of-scope: Report translation, Detection, Report processing, Report data extraction

Excluded: Decision support, Diagnosis

User

Intended: Radiologist, Researcher

Out-of-scope: Layperson

Excluded: Layperson