Research On Classification Of Prognostic Outcome Of SCLC Based On Deep Learning And Radiomics

Small cell lung cancer(SCLC)is a prevalent and high-risk tumor illness with high mortality and morbidity that gravely jeopardizes patients’ health and accounts for around 15-20% of all lung cancers.SCLC is distinguished by invasiveness,fast development,early metastasis,and medication resistance.In clinical diagnosis,the location and segmentation of existing SCLC rely mostly on physician diagnosis by using Computed tomography(CT),which is labor-intensive.Clinical therapy demonstrates that most patients had already suffered severe spread at the diagnosis.However,surgery can only target a relatively small number of patients with restricted stages,resulting in a five-year survival rate for small cell lung cancer of fewer than 5%.How to aid physicians in the exact segmentation of SCLC tumors and investigate efficient markers to assess the therapeutic impact of SCLC is of critical clinical value and relevance for the establishment of the treatment regimen for small cell lung cancer and the clinical diagnosis of doctors.Because of SCLC’s high heterogeneity,intractability,and mortality,the following challenges must be addressed:(1)how to employ computers to help clinicians segment SCLC tumors accurately and quickly in CT images;and(2)how to extract more information from SCLC to better predict the fate of SCLC treatment.To address these issues,this thesis aims to design a deep learning algorithm for segmenting SCLC accurately and to explore whether the fusion of clinical features,radiomics features,and other features of SCLC can be used as a potentially effective marker for classifying the treatment effect of SCLC.The contributions and innovations of this thesis are divided into two main sections.(1)To address the issue of SCLC tumor segmentation relying on manual segmentation,the UL-Net deep learning network is being proposed to aid clinicians in the rapid and accurate segmentation of SCLC tumors in CT images.In this thesis,U-Net is paired with two long short-term memory(LSTM)units to obtain precise tumor segmentation of small cell lung cancer.To begin,the CT images of SCLC are preprocessed to eliminate redundant information and emphasize the tumor characteristics of SCLC;next,two-dimensional planar segmentation is conducted based on U-Net for tumor axial position,and LSTM units are added to restrict the growth of SCLC tumor in coronal location to increase three-dimensional spatial information;and configurable loss functions are used to constrain the UL-Net training;and lastly,a controlled loss function is used to constrain the UL-Net training.The Dice for SCLC tumor segmentation was 0.91±0.04;Recall was 0.96±0.1,and Precision was0.87±0.07.The experiments showed that the UL-Net model could assist radiologists in SCLC tumor segmentation.(2)To address the scarcity of recognized biomarkers for evaluating SCLC stage therapy,a CT radiographic signature-based method is applied for researching validated biomarkers for evaluating SCLC stage treatment.In this thesis,radiomics extracted high-throughput imaging features of SCLC tumors to characterize SCLC tumors prior to staged treatment combined with clinical information from SCLC patients and predicted the effect of staged treatment on SCLC patients based on a combination of deep model features extracted from the pre-trained UL-Net model.Firstly,the CT data was pre-processed,and the region of interest was segmented.Secondly,the radiomics features of SCLC patients’ tumors were extracted using the pyradiomics tool.Thirdly,the depth imaging features were extracted using the middle layer of the pre-trained UL-Net model.Fourthly,the Lasso feature selection was used to select valid radiomics features.The two-sample t-test method was used to select deep image features with significant differences.The U-test method was used to select clinical features with significant differences.Finally,logistic regression(LR)and support vector machine(SVM)were used for each view feature(radiomics features,clinical features,and depth image features)to predict the outcome of the treatment of SCLC.The classification prediction is confirmed by applying the concept of integrated learning to the three fusion classification machine learning models.The AUC values of clinical features,depth imaging features,and radiomics features were 0.67,0.77,and 0.86.Respectively,the AUC value of the multi-feature combination was 0.91.They indicated that radiomics features could be used as effective biomarkers for evaluating the effect of SCLC stage treatment.A multi-feature decision can improve the classification effect of SCLC stage treatment.In conclusion,this thesis proposes a UL-Net model that can effectively segment SCLC tumor regions and assist doctors in accurately segmentation SCLC tumors in CT images.In the meantime,the radiomics features of SCLC tumors mined in this thesis,combined with multi-faceted features,can be used as practical markers to evaluate SCLC stage-specific treatment effects.These findings will act as a basis for future study towards auxiliary clinical diagnosis of SCLC.