Research On Key Issues Of 4DCT Based Radiomics In NSCLC Application

Radiomics is a technique based on high-performance computing and algorithms to extract and analyze features from massive multi-modal medical images such as CT,MRI,and PET/CT.This technique can be used for early diagnosis of disease,patient stratification prediction for personalized treatment and early prediction ot treatment efficacy.The valuable information extracted is provided for the identification of malignant tumors,the management of disease treatment,and individualized precision treatment decision.Currently,the application of radiomics in lung cancer is mainly focused on the following three aspects:first,the identification of t small pulmonary nodules based on low-dose spiral CT and the prediction of benign and malignant with non-invasive methods are explored;second,the utility of radiomics for the specific chemotherapy protocol selection,sensitivity prediction for radiation therapy and anti-angiogenic drugs,selection of targeted drugs,and even sensitivity prediction for immunotherapy is under investigation;third,the models to predict the prognosis of lung tumor based on the characteristics of radiomics signature are developed and validated.However,the accuracy of radiomics analysis technology is based on static high-quality images,while lung tumors are usually affected by the respiratory movements under normal free breathing conditions.Most lung cancer patients have relatively poor lung function who cannot tolerate long-term breath holding state.As such,4-dimensional CT（4DCT）is routinely used for image acquisition for the intolerable patients after simple breathing training.Therefore,based on the aforementioned scenario,the imaging preprocessing,feature extraction and analysis techniques of the radiomics features extracted from lung cancer dataset must be adopted to 4DCT technique,and it is urgently needed to establish a framework of radiomics analysis technical process based on 4DCT situation.The major work in the dissertation include:（1）To investigate the influence of different respiration statuses on the 4DCT radiomics feature extraction.The influence of thirteen image preprocessing methods on the extraction of radiomics festures on 4DCT images was investigate.The results suggest that the working feature of different image preprocessing methods are different.In the case of radiomics analysis,it is necessary to screen out the pretreatment method that meets the analysis requirements according to the purpose of the research,or select those features with good robustness for subsequent analysis under a certain pre-processing method.Meanwhile,we found that under the condition of constant respiratory frequency,with image acquisition procedure twith optimal pitch used,the amplitude of respiratory motion has a postivie correlation with the siginificant difference between the ten phases of 4DCT and static CT.Therefore,the radiomics features extracted from these phases associated with tumor density shows a poor consistency.When the breathing amplitude is the same,the volume difference in exhalation phase images（0%-90%）is smaller than that in the inspiratory phase images（10%-50%）.There was no monotonic correlation between volume difference and respiratory frequency in 4DCT time-division phases and the static CT.In summary,to control the respiratory rate at BPM（Breath per Minute,BPM）=13min^-1 can achieve smaller difference in volume within the 4DCT time-division phases.Different pitches selection have a great influence on 4DCT images.The commonly used optimal pitch selection formula is not suboptimal for 4DCT image radiomic feature extraction because…..Our results show that the optimal pitchfor4DCT should be 0.093.In addition,we found that the maximum end-expiratory EOE（End of the Enpiration,EOE）phase（90%）is relatively less sensitive to pitch variation,with a 70%probability of maintaining a tumor volume deviation of±5%at different pitch settings.（2）To characterize the influence of image preprocessing method on the 4DCT based radiomics analysis.We conducted a detailed study of thirteen image preprocessing methods that may affect the extraction of Radiomics.The results suggest that the Radiomic feature sites of different image preprocessing methods are different,and it is necessary to conduct research based on Radiomics.The purpose is to select a pre-processing method that satisfies the analysis requirements,or select those features with good robustness for subsequent analysis under the pre-processing method that has been completed.This study found Butterworth Smooth-Retest（BSR）,Bit Depth Rescale Range（BDRR）,Laplacian Filter（LF）,Logarithmic Filter（Log Filter,LF）These four image preprocessing methods have the greatest impact on the Radiomic feature extraction.Our experience is to filter out random noise and maintain reasonable image grayscale and resolution as much as possible while preserving the original CT image information.（3）Algorithm for Minimum Gradient Density Projection Reconstruction Image（MGDPM）and its application value in 4DCT based Radiomics research.In order to solve the problem that the radiomic feature extraction of 4DCT images is affected by tumor centroid motion,we innovatively proposed the concept and algorithm of"Minimum Gradient Density Projection Reconstruction Image（MGDPM）",and reconstructed the gradient-weighted 4DCT time-division image.The post-synthesized MGDPM image which is the closest reconstruct 4DCT image to the static CT image is specially used for radiomic feature extraction so that the image information of 4DCT can be utilized to the maximum extent.（4）To explore the efficacy of 4DCT based radiomics analysis 4DCT to predict the distant metasteses in patients with non-small cell lung cancer（NSCLC）finally,we applied the findings to patients with NSCLC.The results demonstrate that the markers based on radiomic features extracted from MGDPM’s images can better predict distant metastases in NSCLC patients,and their predictive power is significanty better than AIP（Average intensity projection,AIP）and MIP（Maximun intensity projection,MIP）reported in previous work.The AUC（Area under the curve,AUC）of MGDPM＿Signature,MIP＿Signature,and AIP＿Signature are 0.826,0.785,and 0.737respectively（p<0.05）.