AI-Based Automated Lipomatous Tumor Segmentation in MR Images: Ensemble Solution to Heterogeneous Data
Chih-Chieh Liu1, Yasser G Abdelhafez2,3, S Paran Yap2
1Department of Biomedical Engineering, University of California, Davis, CA, USA.
Journal of Digital Imaging
|March 1, 2023
Summary
This study developed a deep learning Super Learner (SL) ensemble to accurately segment lipomatous tumors (LTs) despite data variations. The SL framework significantly improved segmentation accuracy, aiding in differential diagnosis of LTs.
Area of Science:
- Radiology
- Medical Imaging
- Artificial Intelligence
Background:
- Accurate segmentation of lipomatous tumors (LTs) is crucial for radiomics analysis and localization.
- Data heterogeneity from varying tumor characteristics and scanning protocols presents a major challenge in LT segmentation.
Purpose of the Study:
- To develop a deep learning-based Super Learner (SL) ensemble framework to improve lipomatous tumor segmentation accuracy.
- To mitigate performance instability and data heterogeneity in lipomatous tumor segmentation using diverse data correction and normalization methods.
Main Methods:
- A Super Learner (SL) ensemble framework was developed using 30 base learners (BLs) trained with fivefold cross-validation.
- Six configurations of data correction and normalization were applied to T1-weighted/proton-density MR images from 185 patients with pathologically proven LTs.
- Performance was evaluated using Dice-Similarity-Coefficient (DSC), sensitivity, specificity, and Hausdorff distance (HD95).
Main Results:
- The SL ensemble achieved an average DSC of 0.80 ± 0.184, outperforming individual base learners.
- SL predictions showed improved sensitivity (0.78 ± 0.193) and specificity (1.00 ± 0.010) compared to BLs.
- The SL significantly reduced Hausdorff distance 95 (HD95) across different tumor locations, indicating enhanced segmentation precision.
Conclusions:
- The proposed deep learning Super Learner framework effectively improves lipomatous tumor segmentation accuracy.
- The method successfully mitigates performance instability and addresses data heterogeneity, aiding in the differential diagnosis of lipomatous tumors in clinical settings.
