Multidimensional Directionality-Enhanced Segmentation via large vision model.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Manual segmentation of retinal fluid in Optical Coherence Tomography (OCT) images is time-consuming and prone to errors.
• Automated methods are needed for accurate diagnosis and treatment planning of macular edema.
• Existing deep learning models, especially Transformer-based ones, struggle with subtle lesions in OCT scans.

Purpose of the Study:

• To develop an automated framework, MD-DERFS, for accurate segmentation of macular edema in OCT images.
• To adapt Transformer-based large vision models for the specific challenges of medical image segmentation in ophthalmology.
• To address limitations in dataset size and annotation scarcity for macular edema segmentation.

Main Methods:

• Introduced the Multidimensional Directionality-Enhanced Retinal Fluid Segmentation (MD-DERFS) framework.
• Developed a Multi-Dimensional Feature Re-Encoder Unit (MFU) for enhanced texture and pathological feature recognition.
• Incorporated a Cross-scale Directional Insight Network (CDIN) for holistic feature analysis and a Harmonic Minutiae Segmentation Equilibrium loss (L_HMSE) for data imbalance.

Main Results:

• MD-DERFS demonstrated superior performance compared to existing segmentation methods on the MacuScan-8k dataset.
• The framework effectively adapts large vision models for boundary-sensitive medical imaging tasks.
• Achieved improved segmentation accuracy for retinal fluid and associated lesions in OCT images.

Conclusions:

• MD-DERFS offers a robust solution for automated macular edema segmentation using OCT imaging.
• The proposed framework enhances the capability of large vision models in specialized medical image analysis.
• This advancement facilitates more efficient and accurate diagnosis and therapeutic planning for macular edema.