Abstract
PURPOSE: Retinal non-perfusion areas (NPAs) are a crucial biomarker for assessing the degree of retinal ischaemia of patients with retinal vascular diseases. Accurate identification of NPA is useful for assessing therapeutic approaches, guiding laser treatments and predicting prognosis. The ultra-widefield swept-source optical coherence tomography angiography (UWF-SS-OCTA) has emerged as a superior non-invasive imaging modality with an outstanding ability to detect NPA. This study aimed to establish an efficient NP-Unet model to accurately segment and quantify NPA in the UWF-SS-OCTA image of branch retinal vein occlusion (BRVO), central RVO (CRVO) and diabetic retinopathy (DR). METHODS: The 29×24 mm retinal images from the posterior pole of patients with retinal vascular diseases (BRVO, CRVO and DR) were acquired by the UWF-SS-OCTA. The enhanced inner retinal OCTA images datasets were constructed for model training. An efficient NP-Unet model was developed to precisely segment and quantify NPA with ischaemic index (ISI). The performance of the NP-Unet model was compared against different typical segmentation models. RESULTS: The NP-Unet model demonstrated its superior capability in segmenting and quantifying NPAs t0 the other models, achieving an accuracy of 93.4%, with an area under the curve of 0.986 and a dice similarity coefficient of 0.904. CONCLUSION: The NP-Unet model can efficiently and accurately segment and quantify NPAs with the ISI in the UWF-SS-OCTA image of BRVO, CRVO and DR, offering a non-invasive, valuable tool for clinical evaluation and treatment guidance.