Transfer Learning Strategy For Medical Image Diagnose

Deep learning makes remarkable progress in image recognition especially convolutional neural networks(CNNs)nowadays,beats other traditional machine learning methods in the first time,wins the champion of Image Net Large Scale Visual Recognition Challenge,and decreases the probability of error of image recognition so much.However,most deep learning model is only used in natural image recognition domain,only a small part is used in medical image diagnose domain.Using deep learning method to make a lung cancer CT image diagnose not only saves a lot of doctors' diagnose time,improves hospitals' working efficiency,solves the insufficient health care resource problem,but also helps to make a diagnose and cure patient earlier even save life.There are two methods of using deep learning method into medical image diagnose:1.Learning from scratch,use medical image to train CNN from scratch 2.Transfer learning,use the pre-trained CNN model architecture and weight parameters to make feature extraction.There is only the learning from scratch method,but nobody uses transfer learning method in lung cancer CT image diagnose now.So we put forward the transfer learning strategy in lung cancer CT image diagnose and do the experiment.The major work:1.Compare and analyze the difference of strategies between learning from scratch and transfer learning in experiment,certify the advantage of transfer learning: avoid the over-fitting problem,reduce a lot of weight parameters and training time,save the computation resource,improve the reuse of models,improves the recognition of lung cancer CT image diagnose.2.Fine-tune the transfer learning model,compare and analyze the advantages and disadvantages of different network architecture models,analyze the reason and condition of transfer learning,use fine-tune strategy and method to get better model and improve the model recognition accuracy.3.Optimize the top model architecture based on the fine-tune model,use a new method of training learning rate to reduce more weight parameters,get a good model convergence result and improve the model recognition accuracy.