Research Of PD-L1 Detection Score Based On Convolutional Neural Network

Cancer is one of the major diseases affecting the health of Chinese residents,and the development of immunotherapy has brought more hope to cancer patients.PD-L1 is a key biomarker for judging whether cancer patients are suitable for immunotherapy,so the quantification of PD-L1 expression has very important research significance in the medical field.The current quantification of PD-L1 expression is assessed by pathologists directly observing under a microscope,but manual scoring is inefficient and highly subjective.In order to solve the above problems,this thesis designed two deep learning methods for the two groups of non-small cell lung cancer and esophageal cancer data to achieve automatic PD-L1 scoring.The specific work is as follows:(1)This thesis strictly follow the latest PD-L1 international working group scoring consensus and doctors’ opinions to allow professional labelers to label the data,build a new dataset.The Gaussian distribution with the cell center as the peak is used to make the label of the cell classification task,so that it can better locate the location of the cell.(2)A combined model algorithm based on ensemble learning is proposed for TPS scoring of non-small cell lung cancer PD-L1 image.The combined model is composed of the Res-unet network and the Link Net network.First,Link Net is used to predict the tumor region segmentation results,and then the predicted tumor region segmentation results and the original image are input into the Res-unet network to predict the cell classification results.The cell classification results are processed and calculated to obtain the TPS score.The final experimental results show that the F1-score value of cell classification can reach 0.7859,which is a certain improvement compared to the original cell classification model.(3)An algorithm based on multi-task learning is proposed for CPS scoring of esophageal cancer PD-L1 images.Based on the Res-unet network and multi-task learning method,the Res-unet network is improved into a multi-task learning network model,so that the network can simultaneously train two tasks of tumor region segmentation and cell classification,and the CPS score can be calculated with the cell classification results.The final comparison verifies that the prediction effect of the multi-task learning method and the combined model method proposed in this thesis is better than the current classical semantic segmentation network.The two deep learning models proposed in this thesis have better prediction effects,and the error of PD-L1 score results is small,which has certain practical significance in assisting doctors in diagnosis and treatment.