Prognostic Analysis Of Non-Small Cell Lung Cancer Based On Quantitative Computed Tomography Image Feature

Imaging-assisted diagnosis approach has become a powerful tool for clinical research and practice.From traditional X-rays to the latest high-resolution multimodal imaging techniques,medical image can provide more diagnostic information.Subtle image heterogeneity could be recognized to describe tumor growth,recurrence and prognosis with the help of current various medical imaging equipment.Based on massive medical image phenotypic feature,the emerging quantitative feature analysis method could capture these subtle image differences.Combined with diagnosis experience of clinicians and statistics analysis algorithms this emerging analysis method could establish a scientific,quantitative,and big data-driven clinical support system.Based on large-scale medical image statistics this approach could provide more real-time and reliable clinical-aided diagnosis than traditional image analysis.Non-small cell lung cancer(NSCLC)has become one of the major cancers that threaten human health in recent years.Its morbidity and mortality rate are increasing year by year.Clinical treatments for NSCLC needs to depend on the diagnose by clinical tissue biopsy.Howerer,the method belongs to invasive treatment.Besides,traditional diagnosis and treatment are difficult to identify the individual difference of patients quantitatively.According to the previous studies,modern quantitative medical image analysis could compensate for this deficiency.This method can accurately predict the clinical prognosis for individual patient in a noninvasive way based on the analysis from thousands of cases,and individualize risk stratification for specific patients according to the patient's individual charactersitics.This paper committed to the quantitative CT image phenotypic feature analysis for individualized NSCLC prognosis.Based on the automatic segmentation and phenotypic feature database construction of NSCLC tumor on CT images,this paper explored the potential relationship between phenotypic features and pathological types and prognosis of NSCLC.Furthermore,a multi-center retrospective study was conducted to slove an actual need of clinical:"how to predict the therapeutic resistant of tyrosine kinase inhibitors(TKI)for stage IV Epidermal Growth Factor Receptor(EGFR)-mutated NSCLC individually".The results proved that the practicability and validity of the proposed image analysis method.The main work of this paper includes the following aspects:An improved toboggan algorithm for automatic detection and segmentation of pulmonary lesions on CT images was proposed.The improved toboggan method was used to search the CT images,and removing the lung parenchyma,blood vessels and other lung tissues automatically.After this approach the lung lesions on CT images were highly preserved.Based on the automatic extraction of lung lesions,an improved 3D region growing algorithm is proposed to segment the lung lesions.Although existing lung lesion segmentation algorithms can implement the task,artificial interaction is generally needed,and it could not satisfy the real-time effective demand of clinical.In addition,the existing lung lesion segmentation algorithms,such as the level set,traditional region growing and graph cut algorithms could not obtain accurate area of lung lesion.Furthermore,for a special kind of lung lesion,ground glass nodule(GGN),segmentation results by the existing methods were undesirable.In this paper our new algorithm used the advantage of togoggan algorithm to identify the region of interest,combined with the intensity difference of pulmonary lesions and lung parenchyma,we put forward a new multi-constrain growing algorithm for lung lesion segmentation.On the other hand,this algorithm could identify different kind of lesions automatically,with different mechanism to segment them,respectively.The results indicated that the methods can accurately locate the lesion region with minor error,and the segmentation speed was ten times faster than the traditional methods.Based on the accurate lung lesion segmentation,a set of CT phenotypic feature of NSCLC was proposed,which including texture,wavelet and other high-dimensional semantic features and clinical experience characteristics(burr,edge,water density and so on).A prediction method for clinical pathology and survial prognosis of NSCLC by quantitative CT image analysis was also proposed.The proposed image feature set consistented with the phenotypic characteristics of NSCLC,which contained first-order statistical features of CT images,texture features,Gabor features and wavelet features,etc.,and a total of 726 phenotypic features.Suppprt Vector Mechine(SVM)was used to learn thousands of NSCLC cases to predict the pathology,TNM staging and survival prognosis of NSCLC.The method was validated on an independent validation dataset.Results represented that the quantitative image feature analysis method could predict the pathology and prognosis of NSCLC accurately for assisting clinicians and guiding clinical decision making.We constructed an individualized therapeutic resistance prediction system to stage IV EGFR-mutant NSCLC who received TKI drugs.According to multi-center pre-therapy CT image analysis,we proposed an image-based prognostic signature that allows risk stratification of TKI drugs in patients with stage IV EGFR-mutant NSCLC.Combined with the signature and clinical risk characteristics,we proposed an individualized prognostic model for TKI therapeutic resistance.This method can be used to evaluate the efficacy of TKI drugs in the treatment of stage IV EGFR mutant NSCLC,and to standardize the clinical management of TKI drugs.