Research And Application Of Lung Tumor Risk Prediction Model Based On CT Image

In order to better detect early lung cancer and effective treatment,it is necessary to regularly screen specific populations.The main technical means is low-dose spiral computed tomography（LDCT）.The emergence of computer-aided diagnosis system can effectively share the growing medical image data analysis and processing work of doctors.With the development and application of artificial intelligence and deep learning,the lung cancer diagnosis system has made great progress,but there are still various restrictive factors that make the system perform poorly in practical application.When giving the diagnosis conclusion,even experienced doctors may be misdiagnosed,let alone relying on the results predicted by the algorithm.Therefore,the computer-aided diagnosis system is in an embarrassing situation.The results obtained by the trust system that medical workers dare not fully release still need to be evaluated and confirmed by themselves,which does not play a substantive role in reducing the workload of doctors.In this regard,we believe that the computer-aided diagnosis system should not only provide a diagnosis conclusion by the system,but also give the corresponding reference value according to the basis of the doctor’s diagnosis.Even if it can not be quantified carefully and accurately and give a general range,it can also help the doctor shorten the film reading time and reduce the workload.Based on this idea,the main research contents and contributions of this paper are as follows:（1）Compared with the existing lung cancer data sets,through the analysis of the feature annotation information given by the data set,six attribute features closely related to the benign and malignant of nodules are selected as the auxiliary diagnosis basis we provide to doctors,and the design idea of multi task model architecture is determined.（2）Compared with the current mainstream deep learning network models,we believe that the multi head attention mechanism in Transformer is conducive to multi task feature extraction from different angles and directions.Therefore,based on the vit model,we have modified it suitable for medical images,and proposed the methods of transforming three-dimensional image data into two-dimensional images suitable for the model Serialize the 3D data to replace the original patch embedding method.（3）Through experimental comparison,we take the encoder part of the vit network transformed by three-dimensional data input as the data sharing layer of our multi task framework,input the extracted features into the subsequent multi task network,and construct a multi task model by using the balanced optimization method of dynamic weighted average in the training process,so as to realize the judgment and output of benign and malignant and various semantic features.（4）Finally,we integrated our multi task model into the complete lung nodule cancer risk prediction system of the research group,received the nodule coordinates output from the previous link,took out the nodule data block,and then used our network to generate diagnostic reports with various nodule attribute characteristics including benign and malignant,which were displayed in the system interface.To realize the whole process integration of computer-aided detection system and computer-aided diagnosis system.