Automated Intracranial Aneurysm Detection and Size Extraction from Imaging Reports Using a Dual-Model Transformer Framework

Abstract Background The unstructured and variable nature of radiology reports remains a barrier to automated extraction of clinically relevant information for large-scale research and AI applications, particularly for intracranial aneurysms (IAs). Objective We aimed to develop and evaluate a dual-model transformer framework for intracranial aneurysm information extraction and assess whether structured processing improved classification and size-extraction performance compared with unstructured text. Methods We analyzed 3,183 neuroimaging reports annotated by neuroradiologists and neurosurgeons. Four encoder-based transformer models were evaluated for aneurysm-presence classification using 5-fold stratified cross-validation. A zero-shot Llama 3.1 framework was used to extract aneurysm size using direct and structured-report prompting. RadDistilBERT and Llama 3.1 were evaluated on 20 temporally distinct reports using full-report and impression-only inputs, with and without structured preprocessing. Results RadDistilBERT achieved the highest mean classification performance, with accuracy of 96.4% (± 1.8%), F1 score of 91.7% (± 10.3%), precision of 94.0% (± 4.6%), and recall of 90.7% (± 15.6%). F1 scores were 87.5% for BERT-base-uncased and 31.1% for ALBERT-base-v2. In the 20-report set, aneurysm-presence classification accuracy was 100% across all input and processing conditions. For aneurysm size extraction, structured report processing increased accuracy from 61.90% to 66.67% using the full report and from 80.95% to 85.00% using the impression section. Conclusions This study demonstrates the feasibility of a dual-model transformer framework for aneurysm-presence classification and size extraction from unstructured neuroimaging reports. RadDistilBERT achieved the strongest classification performance, while Llama 3.1 enabled zero-shot size extraction. These findings support further external validation and expansion to additional aneurysm characteristics with dedicated annotation and validation.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-13
DOI
https://doi.org/10.21203/rs.3.rs-11435195/v1
Primary Topic
Topic Modeling
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Automated Intracranial Aneurysm Detection and Size Extraction from Imaging Reports Using a Dual-Model Transformer Framework

Emily J. Winterhalter, Donald Robert Cantrell, Dillan Prasad, Syed Hasib Akhter Faruqui et al.
1 links
Research Square
Topic Modeling
preprint

Automated Intracranial Aneurysm Detection and Size Extraction from Imaging Reports Using a Dual-Model Transformer Framework

Emily J. Winterhalter, Donald Robert Cantrell, Dillan Prasad, Syed Hasib Akhter Faruqui, Amr R. Alwakeal, Rishi Jain, Matthew Potts, Nikhil Prasad, Michael Oblich
preprint en

Abstract

Abstract Background The unstructured and variable nature of radiology reports remains a barrier to automated extraction of clinically relevant information for large-scale research and AI applications, particularly for intracranial aneurysms (IAs). Objective We aimed to develop and evaluate a dual-model transformer framework for intracranial aneurysm information extraction and assess whether structured processing improved classification and size-extraction performance compared with unstructured text. Methods We analyzed 3,183 neuroimaging reports annotated by neuroradiologists and neurosurgeons. Four encoder-based transformer models were evaluated for aneurysm-presence classification using 5-fold stratified cross-validation. A zero-shot Llama 3.1 framework was used to extract aneurysm size using direct and structured-report prompting. RadDistilBERT and Llama 3.1 were evaluated on 20 temporally distinct reports using full-report and impression-only inputs, with and without structured preprocessing. Results RadDistilBERT achieved the highest mean classification performance, with accuracy of 96.4% (± 1.8%), F1 score of 91.7% (± 10.3%), precision of 94.0% (± 4.6%), and recall of 90.7% (± 15.6%). F1 scores were 87.5% for BERT-base-uncased and 31.1% for ALBERT-base-v2. In the 20-report set, aneurysm-presence classification accuracy was 100% across all input and processing conditions. For aneurysm size extraction, structured report processing increased accuracy from 61.90% to 66.67% using the full report and from 80.95% to 85.00% using the impression section. Conclusions This study demonstrates the feasibility of a dual-model transformer framework for aneurysm-presence classification and size extraction from unstructured neuroimaging reports. RadDistilBERT achieved the strongest classification performance, while Llama 3.1 enabled zero-shot size extraction. These findings support further external validation and expansion to additional aneurysm characteristics with dedicated annotation and validation.

Research Square
Northwestern University (US), Sam Houston State University (US)
Good health and well-being
Topic Modeling
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon