HyFormer: a hybrid transformer-CNN architecture for retinal OCT image segmentation

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Optical coherence tomography (OCT) is crucial for diagnosing and planning treatments for retinal diseases.
• Retinal OCT image segmentation is essential for identifying lesions and tissue structures, aiding ophthalmologists' decisions.
• Accurate segmentation requires networks that capture both global context and fine local details, a challenge due to varying intensities and close proximity of retinal features.

Purpose of the Study:

• To propose HyFormer, an efficient, lightweight, and robust hybrid network architecture for multi-class retinal OCT image segmentation.
• To address the challenge of simultaneously capturing global and local features in retinal OCT images.
• To enhance feature extraction and integration for improved segmentation accuracy.

Main Methods:

• HyFormer utilizes parallel Transformer and convolutional encoders for independent feature capture.
• A multi-scale gated attention block and group positional embedding enhance the Transformer encoder's feature extraction.
• A three-path fusion module in the decoder integrates features, complemented by a class activation map-based cross-entropy loss function.

Main Results:

• HyFormer demonstrated superior segmentation performance and robustness on both private myopic traction maculopathy and public AROI datasets.
• The network effectively segmented retinal layers and lesions associated with age-related degeneration.
• Evaluations confirmed HyFormer's capability in capturing both global and local features for precise OCT image segmentation.

Conclusions:

• HyFormer offers a promising solution for accurate and efficient segmentation of retinal OCT images.
• The hybrid architecture effectively addresses the need for both global and local feature extraction in complex retinal images.
• This approach has the potential to significantly aid in the diagnosis and treatment planning of various retinal diseases.