Pathology Context Recalibration Network for Ocular Disease Recognition

TL;DR

This work tackles ocular disease recognition by integrating pathology context and expert experience priors into deep networks. It introduces Pathology Recalibration Module (PRM) and Expert Prior Guidance Adapter (EPGA) to recalibrate pixel-wise context and highlight informative regions, coupled with an Integrated Loss $L_{IL} = (1-\lambda) L_{CE} + \lambda L_{BS}$ to address class imbalance and sample-wise loss distributions. Empirical results on CASIA2 NC, LAG, and OCTMNIST show PCRNet with IL consistently outperforms state-of-the-art attention-based methods and advanced loss strategies, with visualizations confirming interpretability aligned to clinical diagnosis. The approach demonstrates that incorporating clinical priors enhances both accuracy and explainability, offering practical potential for AI-assisted ocular disease diagnosis.

Abstract

Pathology context and expert experience play significant roles in clinical ocular disease diagnosis. Although deep neural networks (DNNs) have good ocular disease recognition results, they often ignore exploring 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 Recalibration Module (PRM) to leverage the potential of pathology context prior via the combination of the well-designed pixel-wise context compression 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 ocular disease recognition. Additionally, we introduce an Integrated Loss (IL) to boost the ocular disease recognition performance of PCRNet by considering the effects of sample-wise loss distributions and training label frequencies. The extensive experiments on three ocular disease datasets demonstrate the superiority of PCRNet with IL over state-of-the-art attention-based networks and advanced loss methods. Further visualization analysis explains the inherent behavior of PRM and EPGA that affects the decision-making process of DNNs.

Figures (9)

• Figure 1: Top: The flowchart shows how an experienced expert determines the severity level of nuclear cataract (NC) for a subject. First, the expert examines the pathological context of cataracts (e.g., density and location) using the collected AS-OCT images to make a preliminary judgment. Then, drawing on clinical experience, the expert focuses more on the central and lower nuclear regions, which are more closely associated with NC severity. Finally, a conclusive diagnosis is made. Bottom: To mimic this diagnostic process, PCRNet first uses the PRM to generate a coarse context map by leveraging pathological knowledge. Then, the EPGA module produces a pixel-wise attention map to further highlight significant pixel context region guided by expert knowledge. Finally, the augmented feature maps are generated by calibrating the context feature maps using the pixel-wise attention map.
• Figure 2: The framework of the PCRNet with Integrated Loss (IL) is presented to improve ocular disease recognition results based on ophthalmic images. At each stage of PCRNet, we integrate the Pathology Recalibration Module (PRM) and the Expert Prior Guidance Adapter (EPGA) with a residual block to form the Residual-PCR unit. Specifically, the PRM is developed to leverage pathological context prior by combining a pixel-wise context compression operator and pathology distribution concentration operator. Meanwhile, the EPGA further highlights significant pixel-level context regions by effectively incorporating expert experience prior.
• Figure 3: Pathology distribution concentration map ($P$) weight distribution (first column) and expert experience distribution ($E$) (second column) at three stages of PCRNet18 for NC classification: low, middle and high.
• Figure 4: Pathology distribution concentration map ($P$) and expert experience distribution ($E$) statistics of up- (first row) and bottom- (second row) feature map regions based on the PCRNet18 at three stages of a network for NC classification on CASIA2 NC dataset: low, middle, and high.
• Figure 5: Pathology distribution concentration map ($P$) distribution and expert experience ($E$) distribution of PCRNet at three stages for retinal disease classification on OCTMNIST dataset: low, middle and high. The first column denotes pathology context feature weight distribution, and the second column denotes expert experience distribution.