Studies On Radiomics Techniques And Its Clinical Applications

Computer-assisted diagnosis(CAD)can help doctors reduce reading time and improve diagnostic efficiency.As one of the major tools for CAD,radiomics is widely used in clinical research.In this paper,radiomics is applied to the study of many different clinical problems.Several techniques were proposed to improve the modeling process of radiomics based on the nature of the clinical problems and resulted in good results.The main contents of the study include:In radiomics study,it is common to extract a large number of radiomics features to retain the information in the images.However,the number of medical images is often extremely limited.Thus,effective dimension reduction and feature selection are very important to successful radiomics modeling.In this paper,a feature selection method based on scout models were proposed,which can effectively select features that are conducive to modeling.We validated the method using Bra TS2019 and chronic ankle arthritis datasets,and achieved better results than traditional methods.In conventional radiomics studies,radiomics features are usually extracted for analysis in the Region of Interesting(ROI)labeled by the doctor.However,for some specific problems,this approach failed to yield optimal results.To overcome this problem,in the classification problem of lung cancer Epidermal Growth Factor Receptor(EGFR)and other categories,we used the characteristics of PET image to extract a sub-region from the ROI labeled by the doctors.With the features extracted from this sub-region,better classification model was built.With regard to the prognosis of EGFR treatment,we used delta-radiomics,which uses the difference between pretreatment features and post-treatment features,to obtain a better prediction.We also combined automatic segmentation based on deep learning and classical algorithms,radiomics modeling,etc.to establish a fully automatic process for the diagnosis of various clinical problems,including:Ankle stability: a differentiation model based on FS-PD image for ankle stability was established,and the Area Under the Curve(AUC)reached 0.965 in the independent test set.This model can be used to diagnosis ankle stability efficiently.Knee injury: three different models based on cartilage features,subchondral bone features,and T2 values were built for the diagnosis of knee injury.It was found,the features of cartilage lower bone might be a better indicator for knee injury diagnosis than the clinically-used T2 values.This might provide some new vision into the future study on the clinical diagnosis of knee injury.Lung cancer EGFR classification: U-net was used for automatic segmentation of lesions,before radiomics models were built.AUC in the independent test set reached0.903.Cancer metastasis: Mask R-CNN was used to automatically segment the lesions and a classification model to differentiate thyroid cancer metastasis and lung cancer metastasis was established.Three classification model for thyroid cancer brain metastasis,breast cancer brain metastasis and lung cancer brain metastasis was established.