Combination of Active Transfer Learning and Natural Language Processing to Improve Liver Volumetry Using Surrogate
Brett Marinelli1, Martin Kang1, Michael Martini1
1Departments of Radiology (B.M., M.K., M.M., J.T.), Orthopedic Surgery (S.C.), and Neurological Surgery (A.B.C., E.K.O.), Mount Sinai Health System, 1468 Madison Ave, Annenberg Building, 8th Floor, New York, NY 10029; and Department of Medicine, California Pacific Medical Center, San Francisco, Calif (J.R.Z.).
Radiology. Artificial Intelligence
|May 3, 2021
Summary
Weakly supervised learning with active transfer learning and surrogate metrics accelerated deep learning model deployment for liver segmentation. This approach improved accuracy and demonstrated comparable survival prediction to existing methods.
Area of Science:
- Radiology
- Medical Imaging
- Machine Learning in Medicine
Background:
- Deep learning models require extensive data for clinical deployment.
- Liver segmentation is crucial for transplant evaluations and predicting patient outcomes.
- Existing methods for liver segmentation can be time-consuming and may lack precision.
Purpose of the Study:
- To evaluate the efficacy of weakly supervised learning with surrogate metrics and active transfer learning for accelerating deep learning model deployment.
- To improve the accuracy and clinical utility of deep learning-based liver segmentation.
- To assess the performance of deep learning-predicted liver volumes in survival analysis.
Main Methods:
- Leveraged public Liver Tumor Segmentation (LiTS) data and natural language processing of reports.
- Trained a deep learning model for liver segmentation on 239 CT studies.
- Employed an active learning strategy using absolute volume differences to refine model accuracy.
- Compared survival predictions from model-derived liver volumes against radiology reports and MELD-Na scores.
Main Results:
- An active learning-refined model (LiTS-OU) significantly outperformed a model trained solely on public data (231 mL vs. 176 mL absolute volume difference, P = .0005).
- Deep learning-predicted liver volumes demonstrated survival prediction capabilities comparable to traditional radiology reports and MELD-Na scores.
- Active transfer learning enhanced model performance on an institutional test set.
Conclusions:
- Active transfer learning with surrogate metrics effectively facilitated the clinical deployment of deep learning models for liver segmentation.
- The developed model achieved clinically meaningful liver segmentation at a major liver transplant center.
- This approach offers a promising pathway for faster integration of AI in clinical practice.
