每日文献解读

Pressure ulcers remain a persistent challenge in clinical care, particularly for patients with limited mobility, and accurate stage assessment from images is still difficult to achieve in practice. In this work, we investigate how the systematic integration of complementary visual cues can improve automated pressure ulcer staging. All experiments are conducted on a previously released clinical image dataset annotated according to EPUAP (European Pressure Ulcer Advisory Panel) staging standards. Instead of relying on a single modeling strategy, we explore a unified deep learning framework that gradually integrates wound localization, segmentation, and stage classification. Localization is handled using both detection- and segmentation-based approaches, while staging relies on convolutional and attention-based models. The proposed framework performs wound localization automatically and eliminates the need for manual region specification during inference. The results show that segmentation-based localization provides more reliable spatial guidance, with U-Net++ achieving a maximum Dice coefficient of 99.3%. Classification performance improves as global image context, localized wound regions, and attention mechanisms are combined. To provide a more comprehensive performance assessment, all classification models were additionally evaluated using five-fold cross-validation, and statistical significance was assessed using McNemar's exact test. The proposed attention-enhanced two-stream framework achieved 82.64% accuracy on the original train-test split and a mean five-fold cross-validation accuracy of 86.69%. Overall, the findings suggest that pressure ulcer staging benefits from looking beyond a single visual perspective. By examining how localization, feature fusion, and attention interact, this study provides practical insight into building more interpretable and clinically meaningful deep learning systems for automated wound assessment.