Development and validation of a dual-attention deep learning prediction model for noninvasive differentiation of pulmonary invasive mucinous adenocarcinoma from inflammatory pulmonary nodules.

Background: Pulmonary invasive mucinous adenocarcinoma (PIMA) is a rare subtype of lung adenocarcinoma that often mimics benign inflammatory pulmonary nodules (IPNs) on chest computed tomography (CT), leading to diagnostic challenges. We aimed to develop and internally validate a deep learning (DL)-based prediction model for noninvasive differentiation of PIMA from IPN. Methods: This retrospective single-center study included 443 patients with pathologically confirmed PIMA or IPN between January 2021 and December 2023. The reference standard was postoperative histopathology. A total of 1,409 CT slices were used for model development. The dataset was randomly divided at the patient level into training (80%) and validation (20%) cohorts for internal validation. A dual-attention deep learning model (SE-DAS ResNet) integrating spatial and channel-wise attention mechanisms was developed. Model performance was evaluated using the area under the curve (AUC), sensitivity, and specificity with their 95% confidence intervals (CIs), as well as accuracy and F1 score. Results: In the internal validation cohort, the SE-DAS ResNet achieved an AUC of 0.990 (95% CI: 0.978-1.000), an accuracy of 96.6%, a sensitivity of 100% (95% CI: 90.7-100%), and a specificity of 94.1% (95% CI: 84.1-98.4%). The model significantly outperformed a standard ResNet50 baseline (AUC 0.914, P<0.05). An online research platform was developed to facilitate real-time inference. Conclusions: This dual-attention DL prediction model demonstrated excellent performance for differentiating PIMA from IPN on CT images. However, the reliance on an 80/20 internal random split rather than independent external validation may lead to optimistic performance estimates. External validation in independent multicenter cohorts is required to confirm generalizability.