TL-TransUNet: An Improved Lightweight Semantic Segmentation Model of Macular Edema Lesions in Retinal OCT Images

Optical coherence tomography (OCT) offers significant advantages of noncontact operation, high resolution, and real-time imaging, making it particularly suitable for acquiring human retinal images and playing a crucial role in diagnosing and monitoring retinal diseases such as diabetic macular edema (DME). OCT is a key noninvasive imaging modality for retinal diseases such as DME, offering high-resolution visualization of retinal layers and fluid accumulations. However, retinal fluid segmentation faces several challenges including variations in fluid size, location, and shape, as well as complex irregular boundaries. To address these issues, we propose TL-TransUNet, a novel lightweight segmentation model based on TransUNet. The model incorporates a hybrid self-attention mechanism that effectively combines linear self-attention with residual filtered multilayer perceptron modules, reducing both parameter size and computational complexity while capturing global relationships and local details to improve segmentation performance for small lesions. Furthermore, the decoder employs wavelet convolution that utilizes wavelet transform to extract multi-scale features from low- to high-frequency components, enhancing the model's multi-scale learning capability. Experimental results on a public DME dataset demonstrate that our proposed method outperforms several mainstream segmentation approaches, demonstrating superior performance.