Improved segmentation of low-contrast lesions using sigmoid edge model
Amir Hossein Foruzan1, Yen-Wei Chen2
1Department of Biomedical Engineering, Engineering Faculty, Shahed University, Tehran, Iran. a.foruzan@shahed.ac.ir.
International Journal of Computer Assisted Radiology and Surgery
|November 23, 2015
Summary
This study introduces a hybrid algorithm for accurate tumor segmentation in medical images, even with low contrast and noise. The method uses a novel sigmoid edge model for precise boundary detection, improving segmentation robustness for various tumor types.
Area of Science:
- Medical image analysis
- Computational imaging
- Biomedical engineering
Background:
- Traditional image segmentation models assume a step/ramp function for tissue boundaries, which is inaccurate for low-contrast, noisy, or heterogeneous images.
- Partial volume effects and varying tissue textures further challenge accurate tumor boundary identification in medical scans.
- Accurate tumor segmentation is crucial for diagnosis, treatment planning, and monitoring in oncological imaging.
Purpose of the Study:
- To develop a robust hybrid algorithm for segmenting tumors in CT/MR images, particularly those with low contrast and noise.
- To accurately model and delineate tumor boundaries, addressing challenges posed by heterogeneous textures and partial volume effects.
- To improve the accuracy of tumor segmentation by employing a novel sigmoid edge model for boundary refinement.
Main Methods:
- A hybrid approach combining Support Vector Machine (SVM), watershed, and scattered data approximation algorithms for initial tumor segmentation.
- Distinct treatment of small and large abnormalities to optimize segmentation accuracy for different tumor sizes.
- Application of a proposed sigmoid edge model to refine the segmented boundary by accurately fitting the smoothed, noisy intensity profile.
Main Results:
- The algorithm was extensively validated on synthetic and clinical CT/MR datasets, including 57 volumes with diverse tumor characteristics (size, contrast, intensity).
- Achieved a Dice similarity coefficient of [Formula: see text] and an average symmetric surface distance of [Formula: see text] mm on clinical data.
- Demonstrated competitive performance on the IBSR dataset with a Jaccard index of [Formula: see text] and an average runtime of [Formula: see text] s.
Conclusions:
- The proposed hybrid segmentation algorithm, incorporating distinct handling of tumor sizes and a sigmoid edge model, offers a robust solution for diverse tumor types.
- The sigmoid edge model effectively corrects boundaries, enhancing segmentation accuracy in challenging low-contrast and noisy medical images.
- This approach provides a significant advancement in automated tumor segmentation, applicable across various imaging modalities and tumor presentations.
