Automatic Choroid Segmentation and Thickness Measurement Based on Mixed Attention-Guided Multiscale Feature Fusion...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Choroidal thickness variations are key biomarkers for ophthalmic diseases.
• Accurate choroid segmentation in OCT images is vital for clinical diagnosis and monitoring.
• Existing public OCT datasets lack sufficient disease variety and labeled data.

Purpose of the Study:

• To address the challenges in choroidal segmentation due to blurred boundaries, non-uniform textures, and lesions.
• To develop a novel deep learning model for accurate choroidal segmentation and thickness measurement.
• To introduce the Xuzhou Municipal Hospital (XZMH)-Choroid dataset for research.

Main Methods:

• Construction of the XZMH-Choroid dataset with annotated OCT images of normal and eight choroid-related diseases.
• Development of the mixed attention-guided multiscale feature fusion network (MAMFF-Net).
• Integration of a Mixed Attention Encoder (MAE), deformable multiscale feature fusion path (DMFFP), and multiscale pyramid layer aggregation (MPLA) module.

Main Results:

• MAMFF-Net demonstrated superior segmentation performance compared to other deep learning methods (mDice: 97.44, mIoU: 95.11, mAcc: 97.71).
• An automated choroidal thickness measurement algorithm was developed based on MAMFF-Net segmentation.
• Automated measurements closely aligned with the assessments of senior specialists.

Conclusions:

• The proposed MAMFF-Net effectively overcomes challenges in choroidal segmentation from OCT images.
• The developed automated measurement algorithm shows high accuracy, approaching specialist-level performance.
• This work contributes a valuable dataset and a robust model for advancing ophthalmic disease diagnosis and monitoring.