Deep ensemble neural-like P systems for segmentation of central serous chorioretinopathy lesion

Abstract Automatic segmentation of the central serous chorioretinopathy (CSC) lesion and its related ellipsoid zone from the Bruch membrane (EZ-BM) areas is important in the early diagnosis and treatment of retinopathy to prevent vision loss. However, the large variations in the locations and shapes of the CSC lesion, as well as the low contrast of EZ-BM areas with their surroundings make the segmentation task challenging. To address these challenges, in this paper, we propose a new parallel neural-like P system named the deep ensemble neural-like (DEN) P system, which combines the strengths of spiking neural P systems (SN P systems) and deep convolutional neural networks (CNNs) for more accurate and efficient segmentation of CSC lesion and the related EZ-BM areas. The DEN P system establishes three modules with new rules and neuron architectures, which is implemented end-to-end in neurons. Specifically, we propose an ensemble fully convolutional network (FCN) module to train several FCNs with different initializations to obtain effective features, which leverage the strength of ensemble learning in a DEN P system. Benefiting from the parallelism of DEN P systems, FCNs are conducted in different neurons simultaneously. To achieve efficient classification, we propose a multiloss module with three different loss functions to alleviate DEN P system falling into a single-loss function. Different losses are also conducted in different neurons parallelly. To further enhance the performance, we introduce a coarse-fine compensation module to correct detection errors. Being a parallel computational paradigm, DEN P systems are less time consuming, completing CSC lesion and EZ-BM areas on 1,280 images in 0.04 s with an average dice ratio of 0 . 93 ± 0 . 04 and 0 . 95 ± 0 . 02 , respectively. Moreover, the ablation study shows that the proposed modules are critical for effective learning. The extension and generalization of the DEN P systems are also investigated .