Deep Learning With Convolutional Neural Network For Classification Of Small Renal Masses At MSCT:A Preliminary Study

Background: The increasing use of advanced cross-sectional imaging has led to an increase in the diagnosis of incidental small renal masses(SRMs),these SRMs are now becoming a common clinical problem.SRMs are defined as contrast-enhancing masses with a greatest dimension of 4 cm or less on abdominal imaging.Of small renal masses,approximately 80% are renal cell carcinoma(RCC),from a clinical viewpoint,three main RCC subtypes are important: clear cell RCC(ccRCC;80–90%),papillary RCC(pRCC types I and II;10–15%),and chromophobe RCC(chRCC;4–5%).The remaining 20% include benign etiologies,angiomyolipoma(AML)is the most common among them.When a small renal mass is identified incidentally on imaging,the clinical-management challenge involves distinguishing benign masses from those likely to be malignant and determining the appropriate treatment of malignant masses.The traditional approaches for detecting and characterising renal masses are ultrasound(US),computed tomograph(CT),and magnetic resonance imaging(MRI).Most renal masses can be diagnosed accurately using imaging alone.However,CT and MRI features cannot reliably distinguish fat-poor angiomyolipoma(fp-AML)from malignant renal neoplasms and predict the biologic aggressiveness of RCCs.While renal mass biopsy may be considered gold standard for differentiating between renal tumors,it is limited by risk to the patient and additional cost.Furthermore,given the limitations of current conventional imaging options and biopsy technique.A novel and accurate imaging-based method to differentiate malignant from benign renal lesions,and also to distinguish the more aggressive RCC from less aggressive RCC,would be of great clinical value.Deep learning,as a branch of machine learning,which has recently achieved striking performance improvements in diverse fields such as computer vision,image classification,speech recognition,natural language processing and playing games.The convolutional neural network(CNN),is a typical artificial neural network in deep learning,which are composed of several layers of neuron-like computational connections.Convolutional layers of CNN,in which images are processed with several types of filter,are known to be effective for pattern recognition of images.Whereas conventional machine-learning algorithms require features from images to be extracted in advance of learning,application of convolutional layers in deep learning allows the image itself to be used during the learning process.Therefore,deep learning with CNNs enables all of the information contained in the images to be used,though this has been limited according to the explicitly-designed handcrafted feature parameters selected in conventional machine learning,lies at the key advantage.In that way,this technique was quickly adopted by the medical imaging community,it has been proposed and studied in radiology for classification,detection,and segmentation tasks,most of which were published over the past year.Therefore,this method has potential to differentiate renal masses on dynamic contrast agent–enhanced CT images.Section 1 Deep learning with Convolutional Neural Network in Differentiating fp-AML from RCC at MSCTPurpose: To investigate diagnostic performance by using a deep learning method with a convolutional neural network for the differentiation of fp-AML and RCC at dynamic contrast agent–enhanced computed tomography.Materials and Methods: In this institutional review board–approved clinical retrospective study from January 2013 to July 2017,patients with fp-AML and RCC were identified from the pathology database.There were 42 patients with fp-AML(no visible fat at unenhanced CT)and 158 patients with RCC.The 200 cases comprised 124 males and 76 females with a mean age of 55.07 years(range 20–83 years).All patients were examined with a 320-slice dynamic volume CT by using same four phase renal protocol(unenhanced phase,corticomedullary phase,nephrographic phase and excretory phase).The first step of preprocess is the segmentation of lesion site and background removal of each image.Whole-volume renal mass segmentation were performed manually by a board-eligible radiologist for all patients with their information blinded.Unenhanced CT images for each patient were cooperatively reviewed by two radiologists.All consecutive axial CT images in each phase of each tumor were contoured and saved for subsequent process.A total number of 12317 segmented images(there were 786,3854,3900,and 3777 images for unenhanced phase,corticomedullary phase,nephrographic phase and excretory phase respectively)assign for 200 patients were collected as the experiment data.The second step is to resize the images to 96×96 pixels without changing the lesion size.The third step is to place all the images into the train folder containing two folders with two categories,category 0=negative(with all images of fp-AMLs)and category 1=positive(with all images of RCCs).The CNN was composed of four convolutional layers,four maximum pooling layers,a flatten layer and two fully connected layers.This study is composed of two stages: a training stage,in which 10-fold validation was performed and a testing stage to evaluate the performance of the models with new pictures to each model.In the training stage,we emoloyed supervised deep learning with a CNN and two categories as teaching data.Six training processes were carried out with the input of images of each phase respectively,the combination of three enhanced phases,and the combination of all four phases,which perform to establish six different models refer to as model unenhanced,model corticomedullary,model nephrographic,model excretory,model enhanced and model quadruple.In the testing stage,the testing process is simply done by loading the testing image dataset into the classifier.The CNN model returned two numbers respectively representing the possibility of an input image being in each category.The category with the largest possibility will be the decision of the model as an output at last.The diagnostic performance of CNN models in differentiating fp-AML and RCC were evaluated by using a receiver operating characteristic(ROC)analysis.The area under the receiver operating characteristic curve(AUC)for each model was calculated.The mean AUCs of six different models were compared by DeLong’s test.A P value of less than 0.05 was considered to indicate a statistically significant difference.Results: The mean AUC for differentiating fp-AML from RCC by using model unenhanced,model corticomedullar,model nephrographic,model excretory,model enhanced and model quadruple with the test data sets was 0.64(95% confidence interval [CI]: 0.58,0.69),0.83(95% CI: 0.81,0.85),0.83(95% CI: 0.81,0.85),0.81(95% CI: 0.79,0.83),0.85(95% CI: 0.84,0.86),0.84(95% CI: 0.83,0.85).Conclusion: The CNN models except model unenhanced exhibited a high diagnostic performance in distinguishing fp-AML from RCC at dynamic CT.Further research might focus on the application of this tool in other benign renal lesions,such as oncocytomas.Section 2 Deep learning with Convolutional Neural Network for Predicting Fuhrman Grade of ccRCC at MSCTPurpose: To investigate diagnostic performance by using a deep learning method with a convolutional neural network for the prediction of histologic grade in clear cell RCC at dynamic contrast agent–enhanced computed tomography.Materials and Methods: In this institutional review board–approved clinical retrospective study from January 2013 to July 2017,patients with clear cell RCC were identified from the pathology database.There were 69 low grade ccRCC patients and 28 high grade ccRCC patients.The 97 cases comprised 72 males and 25 females with a mean age of 55.64 years(range 28–83 years).All patients were examined with a 320-slice dynamic volume CT by using same four phase renal protocol(unenhanced phase,corticomedullary phase,nephrographic phase and excretory phase).The first and second step of preprocess are in the same method with section one.A total number of 4936 segmented images(there were 390,1499,1547,and 1500 images for unenhanced phase,corticomedullary phase,nephrographic phase and excretory phase respectively)assign for 97 patients were collected as the experiment data.The third step is to place all the images into the train folder containing two folders with two categories,category 0=negative(with all images of low grade ccRCCs)and category 1=positive(with all images of high grade ccRCCs).The CNN was composed of four convolutional layers,four maximum pooling layers,a flatten layer and three fully connected layers.The process used in building models was consistent with section one.The diagnostic performance of CNN models in differentiating grade of ccRCC were evaluated by using a receiver operating characteristic(ROC)analysis.The area under the receiver operating characteristic curve(AUC)for each model was calculated.The mean AUCs of six different models were compared by DeLong’s test.A P value of less than 0.05 was considered to indicate a statistically significant difference.Results: The mean AUC fordistinguishing low-grade ccRCC and high-grade ccRCC inmodel unenhanced,model corticomedullar,model nephrographic,model excretory,model enhanced and model quadruple with the test data cohorts was 0.55(95% CI: 0.49,0.61),0.77(95% CI: 0.74,0.79),0.76(95% CI: 0.74,0.79),0.69(95% CI: 0.66,0.72),0.78(95% CI: 0.76,0.79),0.75(95% CI: 0.74,0.77).Conclusion: Our results show that maybe the CNN models except model unenhanced could be used to differentiate the Fuhrman grade of ccRCC at dynamic CT.This technique requires further validation on a larger scale prior to implementation into clinical practice.