A novel and lightweight Global-context aware image segmentation model

Project Reference :
AISG-100E-2020-055
Institution :
National University of Singapore (NUS), Singapore General Hospital (SGH)
Principal Investigator :
Prof Feng Mengling
Technology Readiness :
3 (Experimental proof of concept)
Technology Categories :
AI - Deep Learning>Medical Image Analysis
Background/Problem Statement
The encoder-decoder model is a popular Deep Neural Network used in medical image segmentation and excels in identifying local patterns when it’s trained with sufficient data. However, it has a limitation and tends to focus too much on local details and misses out on considering the global context of the image, which can be important for handling variations in the data. The lack of global context can lead to inconsistent segmentation performance, particularly in tasks that rely on specific global priors.
The global medical imaging market size is expected to grow to US$ 61 billion in 2030.
Solution
Developed by NUS in partnership with SGH, the innovative deep learning model – Fourier Coefficient Segmentation Network (FCSN), precisely segments objects by learning the complex Fourier coefficients from the object’s masks. This process involves computing the entire object contour. FCSN ensures accurate estimation by integrating the global context of the object, refining segmentation accuracy, and improving its ability to manage unforeseen alterations in medical images, such as noise or blur.
Link to publication:
Global Context Aware Segmentation by Learning the Fourier Coefficients of Objects in Medical Images
Benefits
- When compared to state-of-the-art segmentation models like DeepLab-v3+ and U-Net+ for medical image segmentation tasks, FCSN attained significantly improved performance and achieved a significantly lower Hausdorff score while maintaining a competitive Dice score.
- FCSN is more memory-efficient with the absence of a decoder.
- FCSN’s ability to consider global features ensures consistent performance, unaffected by local variations like contrast changes, noise, or motion blur.
- FCSN is lighter and faster, requiring fewer parameters than existing models while delivering more accurate image segmentations.
Potential Application(s)
FCSN can be used for many practical applications including medical image analysis, computer vision for autonomous vehicles, face recognition and detection, video surveillance, and satellite image analysis.
We welcome interest from the industry for collaboration/ co-development / customisation of the technology into a new product or service. If you have any enquiries or are keen to collaborate, please contact us.