Utility of deep versatile classifier for accurate subtyping of lung adenocarcinoma using hematoxylin-eosin images.

e13700 Background: Subtyping lung adenocarcinoma using histopathology hematoxylin-eosin (H&E) images presents several challenges and can influence clinical decisions. H&E staining alone often proves challenging for definitive subtyping because of the lack of unequivocal morphologic features in some cases. Deep learning models have transformed histopathological image analysis but remain vulnerable to imperceptible perturbations, which pose significant risks in critical clinical applications. Despite advancements in deep learning such as augmentation and regularization, their robustness often lacks accuracy, which highlights the need for a more reliable framework in clinical contexts. Methods: We integrated a deep learning framework for H&E stained images by an attention-based modified ResNet50 architecture to identify unreliable predictions by analyzing their relationship to clinical insights. We used a multitask loss function for precise subtype classification. These refinements enhance model robustness and reliability for histopathological image subtyping. A total of 203,226 patches extracted from 143 H&E hematoxylin and eosin (H&E)-stained formalin-fixed paraffin-embedded (FFPE) whole-slide images (WSI) of non-mucinous adenocarcinoma containing five subtypes: lepidic, acinar, papillary, micropapillary and solid from DHMC repository. These images were labeled according to the consensus opinion of three pathologists and was used to train and test the model through 12,764 training batches. Results: The model has displayed the potential to distinguish subtype prediction of adenocarcinoma from H&E-stained histopathology images with an accuracy of 93.89% (STD: ± 8.82) through 100 epochs. These results, not only denote the capacity of the model to diagnose non-robust samples effectively, but also contrive to remain robust against adversarial vulnerabilities may occur in clinical scenarios. Compared to empirical analysis, this framework demonstrated that improving prediction reliability in high-stakes medical contexts. Conclusions: This study highlights advancements in mitigating vulnerabilities within deep neural network classifiers applied to assist in subtyping of histopathology imaging patches. This advancement enhances clinical decision-making aligned with WHO guidelines by leveraging H&E-stained WSIs to classify invasive non-mucinous adenocarcinomas based on their predominant subtype through patch aggregation.