Related Experiment Video
Updated: Sep 5, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
2.9K
Comparative evaluation of a prototype deep learning algorithm for autosegmentation of normal tissues in head and neck
Jihye Koo1, Jimmy J Caudell2, Kujtim Latifi2
1Department of Radiation Oncology, Moffitt Cancer Center, Tampa, FL, USA; Department of Physics, University of South Florida, FL, USA.
Summary
A new deep learning algorithm for head and neck cancer auto-segmentation shows superior performance compared to commercial software. This validated prototype offers clinically useful results, highlighting the importance of training data in auto-segmentation accuracy.
Area of Science:
- Medical Physics
- Radiotherapy Oncology
- Artificial Intelligence in Medicine
Background:
- Accurate auto-segmentation of organs at risk (OARs) in head and neck (HN) cancer is crucial for effective radiotherapy planning.
- Deep learning (DL) algorithms offer potential for improving the speed and consistency of OAR auto-segmentation.
Purpose of the Study:
- To introduce and validate a novel DL auto-segmentation algorithm for HN OARs.
- To compare the performance of the prototype algorithm against a commercial auto-segmentation software.
Main Methods:
- A prototype DL algorithm utilizing a fully convolutional network (U-Net and V-net) was developed and trained on 864 HN cancer cases.
- Performance was evaluated on 75 validation cases using Dice Similarity Coefficient (DSC), Hausdorff Distance (HD), and Voxel-Penalty Metric (VPM), comparing the prototype (A) with commercial software trained on the same data (B) and different data (C).
- Qualitative assessment of clinical usefulness was also performed on 20 cases.
Main Results:
- The prototype algorithm (A) demonstrated superior performance across all metrics compared to the commercial software (B and C).
- Average DSC/VPM/HD for the prototype were 0.81/84.1/1.6 mm, significantly outperforming commercial software (B: 0.74/62.8/3.2 mm; C: 0.66/46.8/3.3 mm).
- 93% of structures segmented by the prototype were deemed clinically useful.
Conclusions:
- The developed DL prototype algorithm for HN OAR auto-segmentation is validated and shows superior performance.
- Algorithm performance is influenced by the training dataset, emphasizing the need for institution-specific validation.
- The findings support the clinical utility of advanced DL auto-segmentation in radiotherapy planning.
