IntraDxNet: A Deep Learning System with Interpretable Rules for the Intraoperative Differential diagnosis of Bronchiolar Adenoma and Mucinous Adenocarcinoma

Accurately and intraoperatively differentiating between bronchiolar adenoma (BA) and mucinous adenocarcinoma (MA) is critical for preventing overtreatment (e.g., unnecessary lobectomies) or undertreatment (inadequate resections). However, the overlapping histological features found in frozen sections are challenging even for experienced pathologists to address, leading to high diagnostic uncertainty. To address this issue, we developed Intraoperative Diagnosis Network (IntraDxNet), a self-supervised vision transformer (UNI student) with interpretable diagnostic rules (MA/BA ratio thresholds and lesion area criteria) for rapidly and transparently performing classification. When trained on 196 cases acquired from six hospitals and validated on internal (104 cases) and external (27 cases) test sets, IntraDxNet achieved 94.86% patch-level accuracy and superior whole-slide image (WSI) performance (area under the curve (AUC) = 0.94 vs. pathologists’ AUC = 0.75, p  = 0.046) in internal testing cases, reducing the number of uncertain diagnoses by 57.7% while maintaining 100% specificity. The system delivered results within 253 seconds per WSI, satisfying intraoperative time constraints. An external validation (AUC = 0.80) highlighted challenges derived from interinstitutional staining variations but affirmed the regional clinical utility of the model. By embedding rule-based protocols, IntraDxNet eliminated misdiagnoses and prioritized safety, demonstrating its immediate value as a localized diagnostic tool. This study presents the first artificial intelligence (AI) framework that is tailored for intraoperative BA/MA differentiation tasks, bridging innovation with clinical practicality to improve surgical decision-making processes. Future multicenter validation and stain normalization studies are warranted to increase the generalizability of the model.