3D transformer-enhanced multi-scale global co-attention network for precise diabetic macular edema segmentation in OCT volumes

Diabetic macular edema (DME) has emerged as one of the leading causes of visual impairment worldwide, and optical coherence tomography (OCT) plays a pivotal role in detecting DME. Automatic and accurate segmentation of lesions in retinal OCT images is essential for early clinical diagnosis of DME, but most recent deep-learning methods are two-dimensional (2D) segmentation and fail to fully extract the lesions' critical three-dimensional (3D) information contained in OCT images. Here we proposed a novel 3D deep-learning network characterized by combining a transformer encoder with multi-scale feature aggregation, a non-local module, and global channel-spatial joint attention to obtain accurate 3D segmentation of DME and reveal their 3D morphological characteristics. Extensive experimental results demonstrate that our proposed method not only achieves commendable 3D segmentation performance with robust generalization capabilities in challenging cases, but also offers valuable insights into ophthalmic diseases, enhancing the convenience of clinical diagnosis and treatment of DME.