A multi-scale feature fusion network for CNV segmentation in SD-OCT images toward quantitative assessment of neovascular AMD

Choroidal neovascularization (CNV) is a characteristic feature of neovascular age-related macular degeneration (AMD), a leading cause of irreversible visual impairment in the elderly. Accurate segmentation of CNV in spectral-domain optical coherence tomography (SD-OCT) images is crucial for quantitative assessment and treatment monitoring. However, the complex morphology and subtle boundaries of CNV lesions pose significant challenges for fully automated segmentation. In this study, we propose a deep learning-based multi-scale feature fusion network (MFF-Net) for precise and quantitative delineation of CNV in SD-OCT images. The MFF-Net integrates a multi-scale feature fusion module within an encoder-decoder architecture, effectively combining high-level semantic information with low-level details to strengthen skip connections and ensure efficient information flow. To improve boundary delineation, a gradient constraint is incorporated into the loss function, enhancing the network's ability to capture the upper edges of CNV. Furthermore, an optional interactive attention mechanism is introduced to address cases where fully automatic segmentation is insufficient. Ophthalmologists can provide minimal guidance by marking the four extreme points (top, bottom, left, and right) of the CNV region, from which a Gaussian attention heatmap is generated. This heatmap is then concatenated with the original OCT image to guide network retraining, refining the segmentation using interactive prior knowledge. Experimental results demonstrate that MFF-Net achieves reliable CNV segmentation performance. The automatic version of MFF-Net achieves the highest mean Intersection over Union (IoU) score of 0.686 among all compared methods and also outperforms them on other quantitative metrics. Remarkably, when equipped with the interactive attention module, the interactive version of MFF-Net achieves a substantial IoU improvement from 0.686 obtained by the automatic version to 0.802, demonstrating the effectiveness of the proposed interactive mechanism in enhancing segmentation accuracy. These results indicate that the proposed method provides an effective solution for accurate and efficient clinical quantitative assessment.