Pathology Context Recalibration Network for Ocular Disease Recognition

Springer Nature 2021 LATEX template Pathology Context Recalibration Network for Ocular Disease Recognition Zunjie Xiao1, Xiaoqing Zhang1, Risa Higashita1,2* and Jiang Liu1,3,4* 1Research Institute of Trustworthy Autonomous Systems and Department of Computer Science and Engineering, Southern University of Science and Technology, Shenzhen 518055, China. 2Tomey Corporation, Nagoya 4510051, Japan. 3Department of Computer Science and Engineering, Southern University of Science and Technology, Shenzhen 518055, China. 4School of Computer Science, University of Nottingham Ningbo China, Ningbo 315100, China. *Corresponding author(s). E-mail(s): risa@mail.sustech.edu.cn; liuj@sustech.edu.cn; Abstract Pathology context and expert experience play significant roles in clin- ical ocular disease diagnosis. Although deep neural networks (DNNs) have good ocular disease recognition results, they often ignore explor- ing the clinical pathology context and expert experience priors to improve ocular disease recognition performance and decision-making interpretability. To this end, we first develop a novel Pathology Recal- ibration Module (PRM) to leverage the potential of pathology context prior via the combination of the well-designed pixel-wise context com- pression operator and pathology distribution concentration operator; then this paper applies a novel expert prior Guidance Adapter (EPGA) to further highlight significant pixel-wise representation regions by fully mining the expert experience prior. By incorporating PRM and EPGA into the modern DNN, the PCRNet is constructed for automated ocu- lar disease recognition. Additionally, we introduce an Integrated Loss (IL) to boost the ocular disease recognition performance of PCR- Net by considering the effects of sample-wise loss distributions and training label frequencies. The extensive experiments on three ocu- lar disease datasets demonstrate the superiority of PCRNet with IL 1 arXiv:2512.24066v1 [cs.CV] 30 Dec 2025 Springer Nature 2021 LATEX template 2 PCRNet for Ocular Disease Recognition over state-of-the-art attention-based networks and advanced loss meth- ods. Further visualization analysis explains the inherent behavior of PRM and EPGA that affects the decision-making process of DNNs. Keywords: Ocular disease recognition, Pathology recalibration module, Expert prior guidance adapter, Integrated loss, Interpretability 1 Introduction With the global aging population, the eye health has become a pronounced public health concern, mainly caused by visual impairment and blindness [1]. The World Health Organization (WHO) has estimated that approximately 2.2 billion people are suffering from ocular diseases, including cataracts, age- related macular degeneration (AMD), glaucoma, diabetic retinopathy (DR), and myopia [2, 3]. The ophthalmic image-based examination is a commonly used yet effective means for ocular disease screening and diagnosis, which sig- nificantly contributes to detecting these ocular diseases early and reducing the ratio of visual impairment and blindness patients. In clinical practice, clinicians typically make diagnosis conclusions for ocular diseases from ophthalmic images, heavily relying on the pathology con- text and their experience. To be specific, i)Pathology context. They are strongly associated with objective pathology changes of ocular diseases, such as location, shape, and texture features, which can be investigated through ophthalmic images and other medical modalities. ii)Expert experience. It is associated with the extensive clinical training and professional knowledge that clinicians acquired, which also significantly affects the diagnosis conclu- sions. Moreover, Fig. 1 (top) illustrates how an experienced expert gives the nuclear cataract (NC) severity level for a subject based on these two clinical priors from AS-OCT images. First, he investigates the pathology changes (e.g., density and location) of cataracts under the collected AS-OCT images and makes a preliminary judgment according to pathology context. Then, accord- ing to the expert experience, he pays more attention to the central and down nuclear regions, which are more closely related to NC severity. Finally, he made the final diagnosis conclusions. However, effectively infusing pathology context and expert experience into artificial intelligence–assisted diagnosis techniques for automated ocular disease recognition, particularly in deep neural networks (DNNs), has been rarely explored. Over the years, attention mechanism has achieved remarkable success across various learning tasks [4–6], e.g., computer vision and medical image analysis. The critical factor behind the success is that it can enhance the representational capacity of DNNs by guiding them to capture informative context representations. Squeeze-and-excitation (SE) [4] is a representative channel attention method by building long-dependencies among channels. Effi- cient channel attention (ECA) [7] prompts the ide

This content is AI-processed based on open access ArXiv data.