Identification of pathological secretions in bronchoscopic images based on edge features and multi-scale attention mechanism.

The accumulation of pathological bronchial secretions compromises ventilation and oxygenation in critically ill patients and may lead to atelectasis or secondary infection in severe cases, making timely identification and removal of pathological secretions essential during intensive care and surgical anesthesia. Conventional manual bronchoscopic assessment depends heavily on operator experience, lacks real-time reliability, and fails to meet clinical requirements for efficient and precise intervention. To address this limitation, this study proposes an Edge-Aware Dual-Scale Transformer (EADST) for intelligent and automated bronchial secretion recognition based on Canny edge features. Bronchoscopic image data from 50 critically ill pneumonia patients, including both normal physiological and pathological secretions, were preprocessed by grayscale enhancement and Canny edge detection to generate structural representations of secretion regions, which were subsequently processed through a patch embedding module for edge-aware feature mapping, a dual-scale attention module for capturing both global semantic and local structural dependencies, and an edge-aware feed-forward network to adaptively enhance critical channels, followed by a back-end classification head for real-time pathological secretion discrimination. All experiments were conducted under a unified Canny feature representation, and the proposed framework was evaluated against several classic state-of-the-art (SOTA) models, including VGG-16, ResNet-50, EfficientNet, MobileNet, and Vision Transformer (ViT). Experimental results demonstrate that EADST achieves superior accuracy (89.2%) and robustness in pathological secretion recognition on edge-derived features, indicating that attention-driven and edge-adaptive feature modeling effectively enhances bronchoscopic visual perception and providing a promising foundation for intelligent, real-time bronchoscopic pathological secretion aspiration decision support systems.