Research On Lesion Region Segmentation And Disease Classification Of Lung Tumor Pathological Tissue Image Based On Deep Learning

Targeted therapy is an effective treatment method for non-small cell lung cancer.Different types of lung cancer tumors require different molecular targeted drugs.Therefore,before treatment,pathologists need to confirm the tumor type by observing the patient’s pathological tissue sections,which is a very time-consuming and highly repetitive work.In this study,a deep learning-based method for lung cancer pathological tissue image classification and segmentation was proposed,constructing a multi-task convolutional neural network to simultaneously realize tumor lesion area segmentation and histological subtype classification.Pathologists can quickly locate the lesion area through the model and quickly complete the pathological analysis report based on the histological subtyping pre-classification results provided by the model,ultimately providing strong support for the clinical program formulation of lung cancer molecular targeted therapy.The specific research of this paper is as follows:This paper uses the lung cancer pathological tissue image dataset provided by the pathology department of Qilu Hospital of Shandong University to propose a new end-to-end multi-task convolutional neural network.It uses a shared feature extraction channel to simultaneously provide multi-scale features for segmentation and classification tasks,and then synchronously realizes lesion area segmentation and histological subtype classification in a dual-branch manner.At the connection between the shared feature channel and the dual-branch structure,we use a special receptive field design to ensure that the model can adapt to input images of any size,expanding the practical application range of the model.Secondly,the competition for shared channel parameter updates caused by dual-branch structures during backpropagation leads to a new weighted loss function proposed to balance the competition between segmentation and classification tasks for shared channel parameter updates.By adjusting the loss weight ratio,the model’s performance optimization is achieved,and the dual-branch competition mechanism and impact are discussed.Finally,we evaluated the effectiveness of our model on segmentation and classification tasks and conducted a detailed analysis.The proposed model in this study can accurately segment and classify input data,especially in terms of segmentation performance,which can more accurately and delicately identify and segment lesion areas than manual labeling.In comparison experiments,our method is superior to other methods that can simultaneously achieve medical image segmentation and classification and performs well in lesion area segmentation and histological subtype classification performance.