Artificial Intelligence Assisted Diagnosis Based On Convolutional Neural Networks For Bone Metastasis Of Malignant Tumor By Bone Scintigraphy And CT Imaging

Bone scintigraphy or bone scan（BS）imaging with ^99mTc labeled methylene diphosphonate（MDP）as imaging agent and computed tomography（CT）are the most commonly used two imaging methods in clinical diagnosis and screening of bone metastasis of malignant tumor.However,in the interpretation work of CT and BS imaging for the diagnosis of bone metastasis of malignant tumor,the problems such as excessive workload,repeated work,missed diagnosis and misdiagnosis always exist even in experienced doctors.Recent studies have found that the computer-aided diagnostic system（CADS）of medical imaging based on artificial intelligence（AI）has promising clinical application value in reducing the workload of doctors and reducing the missed diagnosis rate of lesions.CADS with high diagnostic accuracy may solve some clinical difficulties in the future.Objective:The purpose of the study was to construct a computer-aided diagnostic system（CADS）and a computer-aided detective system（CADe S）using the convolutional neural networks（CNN）in deep learning（DL）to diagnose and detect bone metastasis of malignant tumor automatedly.CADS and CADe S were based on the two most commonly used imaging methods in screening and diagnosing bone metastasis of malignant tumor,^99mTc-MDP whole-body bone scan imaging and computed tomography respectively.The diagnostic or detective ability of the two AI model would be evaluated and their influencing factors of results would be analyzed.Materials and methods:The BS images of patients with malignant tumor who underwent ^99mTc-MDP whole-body bone scan from January 2016 to March 2019 in the department of nuclear medicine,West China Hospital of Sichuan University were included as the training,validating and testing data sets of CADS-1 model.Image Net＿Resnet-50network would be used to construct CADS-1.The diagnostic sensitivity,specificity,accuracy,positive predictive value（PPV）,negative predictive value（NPV）and area under the curve（AUC）of receiver operating characteristic（ROC）curve were used to evaluate the diagnostic performance of CADS-1.The included BS images were grouped according to age,primary tumor type,number of lesions and extent of disease（EOD）.The diagnostic results of CADS-1 among different groups were compared to analyze the factors influenced the diagnostic performance of CADS-1.The diagnostic ability of CADS-1 and physicians was compared by comparing the diagnostic results and time consuming of CADS-1 and three physicians in diagnosing the same BS image set.The study also evaluated whether the diagnostic results of CADS-1 can help doctors to make diagnostic decisions by comparing the diagnostic accuracies of doctors after and before referring to the diagnostic results of CADS-1.The BS images of prostate cancer（PC）patients who were underwent^99mTc-MDP whole-body bone scan before March 2019 in our department were included as the training,validating and testing data sets of CADS-2 model.Image Net＿Resnet-50 network would be used to construct CADS-2 also.The difference of the sensitivity,specificity,accuracy,PPV,NPV and AUC of ROC curve between CADS-1 and CADS-2 would be compared to discuss whether disease specific CADS-2 of prostate cancer is more valuable than the comprehensive system CADS-1 based on multi disease bone imaging image in diagnosing bone metastasis of prostate cancer.The images of patients with malignant tumor who underwent CT examination in our hospital before March 2019 were included as the training,validating and testing data sets of the CADe S model.Deep Lab V3+would be used as the basic segmentation network.Res Netv2-50 would be used as the backbone network in the coding stage.The pyramid spatial pyramid pooling（ASPP）module would be used for multi-scale feature extraction and fusion to capture multi-scale context information.The sensitivity and positive predictive value of detection would be used to evaluate the detective ability of CADe S for osteogenic bone metastases on chest CT.The detective sensitivity and PPV of CADe S for metastases at different bone sites would be compared and the causes of false positive（FP）and false negative（FN）detection would be analyzed.Results:In our study,a total of 12222 whole-body BS images of patients with malignant tumors were included to construct CADS-1.The included images were randomly assigned to training set（n=9976）,validating set（n=1223）and testing set（n=1223）.The diagnostic sensitivity,specificity,accuracy,PPV and NPV of CAD-1 in its testing set images were 92.64%,93.92%,93.38%,91.75%and 94.59%respectively.The AUC of the ROC curve was 0.964（95%CI:0.952-0.976）.When the testing set images were divided into groups according to the number of positive lesions,the difference of diagnostic results of CADS-1 among each group was significant.The diagnostic accuracy of CADS-1 for bone scan images with 0,7-10 and more than 10 lesions were 98.60%,98.51%and 98.63%respectively,which were pretty high.The diagnostic accuracy of CADS-1 for bone scan images with 1,2-3 and 4-6 lesions were 82.61%,90.35%and 94.17%respectively,which were lower than the previous three groups.When the testing set images were divided into groups according to EOD,the difference of diagnostic results of CADS-1 among each group was also significant.CADS-1 has the highest diagnostic accuracy（95.88%）of bone scan images with EOD of 3 bone sites,which was significantly higher than images with EOD of 1 or 2 bone sites（85.58%and 92.02%）.However,when the testing set images were divided into groups according to the patient’s age,primary tumor type and lesion location,there was no significant difference exists in the diagnostic results of CADS-1 for each group.Without reference to other clinical data of patients,the highest and average diagnostic accuracy of three doctors for BS images were 89.00%and 85.75%respectively,which were significantly lower than 93.50%of CADS-1（P<0.05）.The diagnostic accuracy was improved when doctors referred to the diagnosis results of CADS-1,but the difference was not statistically significant（P>0.05）.It averagely took 136.33 minutes for three doctors to diagnose 400 bone scan images,and only11.3 seconds for CADS-1 to diagnose the same images,saving 99.89%of the time.A total of 4048 bone scan images of PC patients were included in the study.Randomly assigned included images to training set（n=2833）,validating set（n=405）and testing set（n=810）.The diagnostic sensitivity,specificity,accuracy,PPV and NPV of CADS-2 in testing group were 85.52%,95.41%,90.99%,93.46%and89.37%respectively.The AUC of ROC curve was 0.940（95%CI:0.844-0946）.In the diagnosis of 240 new BS images of PC patients,the specificity of CADS-2 was higher than that of CADS-1（85.00%vs.93.33%,χ²=4.313,P=0.038）,and the AUC of CADS-2（0.941（95%CI:0.910-0.972））was larger than that of CADS-1（0.870（95%CI:0.821-0.919））.A total of 160 CT images with osteogenic bone metastatic lesion were included.and 887 lesions in these 160 CT images were manually marked by doctors.The 128CT images with 733 metastatic lesions were randomly assigned to the training set and 32 CT images with 154 metastatic lesions were randomly assigned to the testing set.The sensitivity and PPV of detection for metastatic lesion on testing set images were 85.71%and 83.02%respectively.The sensitivity of CADe S for detecting skeletal metastatic lesions for thoracic vertebrae,ribs,scapula,sternum and humerus were 91.14%,84.74%,60.00%,66.67%and 100.00%respectively（χ²=7.893,P=0.096）.A total of 27 false positive lesions and 21 false negative lesions were detected by CADe S.The most common cause of false positive detection is skeletal degeneration.The main reason of false negative detection is small size and low density of lesions.Conclusion:Based on the most commonly used two imaging methods for the diagnosis of bone metastasis of malignant tumor,a computer-aided diagnosis and a computer-aided detection system were successfully constructed using CNNs,which realized the automatic diagnosis and detection of bone metastasis of malignant tumor.The constructed systems obtained high sensitivity,specificity,accuracy and diagnostic value.In the study with the current data volume of artificial intelligence assisted diagnosis of bone metastasis of malignant tumor based the whole-body bone scan imaging,the diagnostic performance of the disease-specific model CADS-2 for prostate cancer patients was better than the comprehensive model CADS-1 trained by multi disease images.The number of positive lesions and the extend of disease are the two main factors that can influence the diagnostic results of the AI system based on bone scan imaging.Lesion size and lesion density are the two main factors that can influence the detective results of the AI system for osteogenic bone metastases based on CT images.Bone degeneration is the most common cause of false-positive detection.