Simultaneous classification and delineation of seven types of histological growth patterns in lung adenocarcinomas using self-supervised learning and online hard patch mining.

Background: According to the international standard, identifying and delineating the various growth types of adenocarcinomas is essential to the grading of lung cancer, the world-leading cause of cancer-related death. However, the task is currently conducted by manual inspection of the whole-slide images (WSIs), which is time-consuming and depends on individual experience. The aim of this study is to develop a patch-based deep learning (DL) framework that accurately classifies and delineates lung adenocarcinoma (LUAD) growth patterns in WSIs, addressing the time-consuming and experience-dependent nature of manual inspection while achieving performance comparable to experienced pathologists. Methods: This study presents a patch-based DL framework for classifying and delineating normal lung tissue and seven LUAD growth patterns in WSIs. To boost model performance, we combine self-supervised learning with consistency regularization and pseudo-labeling, using both labeled and unlabeled data. Our pseudo-labeling strategy applies weak and strong data augmentations to generate high-confidence pseudo-labels for unlabeled patches, assigning them only when predictions exceed a confidence threshold to ensure reliability. Additionally, we introduce an online hard patch mining method to improve feature extraction for histologically similar patches that are challenging to classify during training. Results: The model was trained and evaluated on a dataset of 288 WSIs, encompassing seven types of LUADs and normal tissues, achieving a precision of 95.3%±5.5% and a recall of 96.8%±2.9%, with an overall Intersection over Union (IoU) of 87.5%±6.9%. When benchmarked against other state-of-the-art models in LUAD classification using a publicly available dataset, our model demonstrated superior performance. For example, compared to DiPalma Conclusions: The experimental findings demonstrate that the proposed framework achieves high accuracy in classifying normal lung tissues and seven types of lung tumor growth patterns, with its performance across all evaluation metrics rivaling that of pathologists' assessments. This model is capable of generating predictive probability maps at the level of full-slice lung images. These maps effectively visualize the location, distribution, and proportion of lesion areas, thereby facilitating pathologists in conducting further grading and staging tasks. Moreover, when the model's classification results are integrated with pathologists' diagnostic outcomes, the accuracy and reliability of the pathologists' clinical interpretations can be significantly enhanced. Besides, Hematoxylin and eosin staining and standardization measures ensure the method's generalizability across institutions or scanners.