Diagnosis system for retinopathy of prematurity with Fourier parameterized rotation equivariant convolutions network and prompt mechanism

Abstract

Retinopathy of Prematurity (ROP) represents a significant ophthalmic disorder in preterm infants, posing substantial risks to visual development. While deep learning-based approaches have been increasingly applied to ROP diagnosis, current research predominantly focuses on plus disease detection and basic screening/staging of ROP, with insufficient attention to the critical aspect of disease zoning. Moreover, the integration of automated staging and zoning, which is essential for comprehensive disease severity assessment, remains unexplored in existing literature. Against this background, we propose a novel dual-task neural network framework, SegClass-Net, which incorporates Fourier series expansion-based equivariant convolution (F-Conv) for simultaneous segmentation and classification tasks. This framework is specifically designed to perform precise segmentation of the optic disc (OD) and retinal lesions while concurrently generating diagnostic outputs encompassing both staging and zoning information. The methodological innovation lies in the implementation of F-Conv, which significantly enhances segmentation precision through its advanced feature extraction capabilities. Furthermore, we introduce a novel prompting mechanism that utilizes lesion segmentation results as prior information to refine staging accuracy. This integrated approach not only establishes a foundation for accurate ROP zoning but also enhances overall diagnostic performance through synergistic information utilization. Extensive experimental evaluations demonstrate the effectiveness of our approach, with segmentation precision reaching 96.00 % for OD and 90.81 % for lesions, respectively. Notably, the overall ROP diagnostic accuracy achieves 91.78 %, representing a 6.85 % improvement over conventional methods that treat staging and zoning as separate tasks. These results suggest that SegClass-Net offers a promising solution for comprehensive ROP assessment, potentially facilitating earlier intervention and improved clinical outcomes in neonatal ophthalmology.