Application Of Deep Learning In Lung Cancer Pathological Image Classification

Lung cancer is the malignant tumor with the highest morbidity and mortality in China,among which lung adenocarcinoma and lung squamous cell carcinoma are the subtypes with the highest proportion.Histopathological examination is the "gold standard" for the diagnosis of lung cancer subtypes.Traditional pathological diagnosis involves pathologists screening whole slide images to identify tiny lesions from the complex tumor microenvironment to give diagnostic results.This process suffers from time-consuming,low accuracy and poor consistency,which cannot meet the needs of patients.The development of smart medicine and computer-aided diagnosis provides new ideas for the diagnosis of lung cancer subtypes.Intelligent learning and analysis of lung cancer pathological images through deep neural networks can accurately and efficiently assist doctors in diagnosis.The main research goal of this paper is to construct a high-precision deep learning model to achieve fast and accurate lung cancer subtype diagnosis.The main work and innovation points of this paper are summarized as follows:(1)Aiming at the problem of obvious color difference in whole slide images of lung cancer,we improve the structure preserving and sparse stain separation color normalization algorithm which can effectively remove color artifacts and make different pathological images rendered with the same color distribution with maximum preservation of cells,tissues and morphological structures.(2)We proposed a fully supervised lung cancer pathology image classification algorithm based on decision level fusion.Firstly,we extract the lesion area of whole images and crop them into small patches for the learning of the network.The decision-level fusion of the network’s prediction results for small patches using the majority voting method obtains the prediction results for the whole slide image,which effectively equalizing the classification bias of the patch-level images and improving classification accuracy and robustness.In this paper,a dataset of pathology images containing lesion annotations is constructed and experimented on different classification tasks,which demonstrates that the model has extremely high classification performance for both patch-level images and whole slide images.(3)We succeeded a lung cancer pathology image classification algorithm based on multiple example learning and attention mechanism.To address the problem that the pathology image pre-processing requires a lot of manual annotation,we developed a standard pathology image pre-processing toolbox to screen out the white background,natural holes and glial areas and achieve automatic cropping of histopathological areas.Besides,we designed a deep-gated attention model integrating two different activation function features to achieve feature aggregation for the patches.In addition,we introduce an additional binary clustering algorithm as an aid to constrain the patch-level feature space so that the positive and negative features in each category are as linearly separable as possible.Finally,a high-resolution,high-fine-grained attention heatmap based on the attention score is generated which can reveal the probability of cancer while preserving the underlying cellular,tissue and morphological structures.In this paper,we construct a large-scale pathological image dataset for experiments and demonstrated that the model is able to learn diagnosis-relevant morphological features from the tissue microenvironment without any prior knowledge or fine-grained annotation,and achieve the correct classification of lung cancer subtypes and the accurate detection of lesion regions.The accuracy of the proposed algorithm for lung cancer subtype diagnosis was demonstrated to be higher than that of other multiple instance learning algorithms.