Pixel-Level Classification Of Histologic Patterns Of Lung Adenocarcinoma

Lung adenocarcinoma is the most common subtype of lung cancer,and its pathological diagnosis is crucial for treatment.According to the latest tumor classification criteria published by the World Health Organization,invasive non-mucinous lung adenocarcinoma can be divided into five histological patterns,including lepidic,acinar,papillary,micropapillary,and solid patterns.Of these,lepidic predominant tumors have the best prognosis,acinar or papillary predominant tumors have an intermediate prognosis,and micropapillary or solid predominant tumors have the worst prognosis.Therefore,accurate classification of these patterns is of great significance for clinical diagnosis and prognostic treatment.In this paper,we propose using deep learning to perform pixel-level classification of the five histological patterns of lung adenocarcinoma,which can assist pathologists in determining tumor grading and improve diagnosis efficiency.Compared to other existing methods,we have improved the classification accuracy from regional to pixel-level classification,providing more detailed information for clinical diagnosis.Based on the pathological characteristics of lung adenocarcinoma and the clinical guidance of pathologists,we manually annotated a dataset consisting of 1000 images of512×512 pixels(200 images for each pattern),420 images of 1024×1024 pixels,and 4images of whole slide imaging(WSI).In order to obtain more datasets,we developed a data stitching method to augment the training dataset.We proposed an analysis framework for lung adenocarcinoma pathology images that can intelligently annotate the histological patterns at the pixel level.The framework consists of five deep neural network model branches for segmenting the different patterns.We demonstrated the feasibility of our data stitching method through experiments and showed that our trained model outperformed other methods in terms of Dice similarity coefficient(DSC)by 24.06% on a WSI.Through comparison experiments with other networks(U-Net,LinkNet,FPN),our network model outperforms the DSC score of the comparison network by up to 10.78% in a test set containing 200 images.Meanwhile,the overall accuracy of prediction in four WSIs is 99.6%.The proposed network model shows better accuracy in classifying lung adenocarcinoma histological patterns at pixel level,and demonstrates better accuracy and robustness in most cases compared to other network structures.