Weakly supervised detection of central serous chorioretinopathy based on local binary patterns and discrete wavelet...

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Vision

Background:

• Central serous chorioretinopathy (CSCR) is a prevalent fundus disease.
• Early detection is crucial for preventing vision loss in CSCR patients.
• Existing detection methods can be labor-intensive and require precise lesion segmentation.

Purpose of the Study:

• To develop and validate a novel, automated method for detecting CSCR.
• To improve the efficiency and accuracy of CSCR diagnosis using advanced image processing and machine learning techniques.
• To avoid the need for manual segmentation of CSCR lesions.

Main Methods:

• Integration of discrete wavelet transform (DWT) for image decomposition and local binary patterns (LBP) for robust texture feature extraction.
• Application of multi-instance learning (MIL) to bypass the need for precise lesion localization and segmentation.
• Utilizing high-frequency components from DWT for detailed feature extraction, minimizing image interference.

Main Results:

• The proposed method achieved a high accuracy of 99.58% with a single threshold (K=35) using a high-frequency feature fusion scheme.
• Further enhancements through multi-threshold optimization (MTO) and integrated decision-making (IDM) improved detection accuracy to 100% (K=40).
• Experiments were conducted on a dataset of 358 optical coherence tomography (OCT) B-scan images, demonstrating significant effectiveness.

Conclusions:

• The developed automated method effectively detects central serous chorioretinopathy (CSCR).
• The integration of DWT, LBP, and MIL offers a robust and accurate approach for CSCR diagnosis.
• This technique shows competitive performance compared to existing methods and holds promise for clinical application.