Quadratic divergence regularized SVM for optic disc segmentation.

Area of Science:

• Ophthalmology
• Computer Science
• Machine Learning

Background:

• Machine learning models for retinal image processing, like optic disc segmentation, assume consistent data distributions.
• Retinal images often exhibit varying feature distributions due to different collection conditions, leading to poor model performance across datasets.
• Relabeling data and retraining models for each new dataset is costly and time-consuming.

Purpose of the Study:

• To develop a novel transfer learning method to address domain shift in retinal image analysis.
• To enable effective optic disc segmentation even with limited or no labeled data in the target domain.
• To propose a quadratic divergence regularized support vector machine (QDSVM) for knowledge transfer.

Main Methods:

• The proposed QDSVM method minimizes the distribution difference between source and target domains during classifier training.
• Simultaneously trains a classifier while aligning feature distributions.
• Applies transfer learning to overcome domain shift challenges in medical image analysis.

Main Results:

• The QDSVM method significantly reduced superpixel-level classification error from 14.2% (without transfer learning) to 2.4% (with transfer learning).
• Demonstrated effective knowledge transfer from source domains with abundant data to target domains with limited or no data.
• Validated the method's efficacy for optic disc segmentation across datasets with differing feature distributions.

Conclusions:

• The proposed QDSVM is an effective transfer learning approach for retinal image segmentation.
• It successfully bridges the domain gap, improving model generalizability.
• Enables robust optic disc segmentation without extensive data relabeling for new datasets.