Related Experiment Video
Updated: Sep 29, 2025
07:15
Machine Learning Algorithms for Early Detection of Bone Metastases in an Experimental Rat Model
Published on: August 16, 2020
7.0K
Developing a Cancer Digital Twin: Supervised Metastases Detection From Consecutive Structured Radiology Reports.
Karen E Batch1, Jianwei Yue1, Alex Darcovich1
1School of Computing, Queen's University, Kingston, ON, Canada.
Frontiers in Artificial Intelligence
|March 21, 2022
Summary
Natural language processing (NLP) improves metastatic cancer detection by analyzing past radiology reports. This approach enhances prediction models, enabling better tracking of cancer progression over time.
Area of Science:
- Oncology
- Radiology
- Artificial Intelligence
- Natural Language Processing
Background:
- Digital cancer twins require high-resolution patient data for treatment monitoring.
- Accurate detection of metastatic disease over time is crucial for patient management.
- Current methods for analyzing radiology reports can be time-consuming and may miss subtle progression patterns.
Purpose of the Study:
- To enhance the detection of metastatic disease over time using historical patient data from radiology reports.
- To investigate the efficacy of Natural Language Processing (NLP) in generating weak labels for semi-supervised classification of computed tomography (CT) reports.
- To compare the performance of different machine learning models in predicting metastatic disease based on sequential radiology reports.
Main Methods:
- Utilized 714,454 structured radiology reports from Memorial Sloan Kettering Cancer Center.
- Developed and validated NLP models, including Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs), on curated datasets for lung, liver, and adrenal metastases.
- Extracted and encoded features from consecutive radiology reports to train models for predicting metastatic disease, comparing them against a single-report baseline model.
Main Results:
- NLP models exposed to consecutive reports significantly improved the classification of metastatic disease compared to single-report models.
- The best-performing model achieved higher accuracy, precision, recall, and F1-score in predicting metastases.
- The study successfully generated metastases maps for over 714,454 reports, demonstrating an automated labeling approach.
Conclusions:
- NLP models can effectively extract cancer progression patterns from multiple sequential radiology reports.
- Analyzing historical report data with NLP offers a more performant and automated method for detecting metastatic disease.
- This approach provides a time- and cost-effective solution for tracking cancer progression and supports the development of digital cancer twins.
