Related Experiment Video
Updated: Sep 29, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
3.0K
Deep Learning and Domain-Specific Knowledge to Segment the Liver from Synthetic Dual Energy CT Iodine Scans
Usman Mahmood1, David D B Bates2, Yusuf E Erdi1
1Department of Medical Physics, Memorial Sloan Kettering Cancer Center, New York, NY 10065, USA.
Diagnostics (Basel, Switzerland)
|March 25, 2022
Summary
Deep learning maps single energy CT (SECT) scans to synthetic dual-energy CT (synth-DECT) iodine scans. This improves liver segmentation accuracy, requiring less training data for better performance.
Area of Science:
- Radiology
- Medical Imaging
- Artificial Intelligence
Background:
- Single energy CT (SECT) scans are widely used but lack material density information.
- Dual-energy CT (DECT) provides material density information, crucial for applications like iodine quantification.
- Liver segmentation is vital for treatment planning and monitoring in various liver diseases.
Purpose of the Study:
- To develop a deep learning (DL) model to generate synthetic DECT (synth-DECT) material density iodine (MDI) scans from SECT scans.
- To evaluate the effectiveness of synth-DECT scans for improving liver segmentation accuracy compared to SECT scans.
- To assess the data efficiency of DL models trained with synth-DECT versus SECT data for liver segmentation.
Main Methods:
- A 2D pix2pix (P2P) deep learning network was trained on 100 abdominal DECT scans to map SECT to synth-DECT MDI scans.
- The P2P algorithm transformed 140 public SECT scans into synth-DECT scans.
- Four existing liver segmentation frameworks were trained and tested using both synth-DECT and SECT scans, with accuracy measured by Dice Similarity Coefficient (DSC).
Main Results:
- Synth-DECT-trained models achieved higher average DSC scores (0.93±0.06 on held-out, 0.89±0.01 on generalization sets) compared to SECT-trained models (0.89±0.01 and 0.81±0.02, respectively).
- Models trained with synth-DECT data achieved comparable or better performance than SECT-trained models while requiring less training data.
- Improved segmentation accuracy was observed across both held-out and generalization test sets when using synth-DECT scans.
Conclusions:
- Deep learning-based mapping of SECT to synth-DECT MDI scans is a valuable technique for enhancing liver segmentation.
- Synth-DECT scans improve segmentation accuracy and reduce the amount of training data needed for DL models.
- This approach holds promise for improving the efficiency and effectiveness of AI-driven medical image analysis in radiology.
