CT Based Deep Learning For The Evaluation Of Acute Pancreatitis:Diagnosis,Severity Prediction And Pancreatic Infection Assessment

Objective:To developed and validated a computed tomography（CT）based deep learning（DL）model for the diagnosis of acute pancreatitis（AP）,the prediction of severity grading under the Revised Atlanta Classification,and the assessment of secondary pancreatic infection（PI）.Materials and Methods:Chapter 1:Nine hundred and ninety AP patients admitted to the Department of Integrated Traditional Chinese and Western Medicine,West China Hospital of Sichuan University from July 2016 to August 2019 and underwent non-enhanced computed tomography（NECT）scan within 72 hours of disease onset were retrospectively included.Meanwhile,416 non-acute pancreatitis（NAP）people from physical examination center of our hospital from January 2017 to August 2019 were retrospectively collected as the normal control.Patients were divided to a training dataset（782 AP and 343 NAP）and an independent testing dataset（208 AP and 73NAP）.A DL system based on the lightweight network（modified 3D Mobile Net V2）was proposed to process 3D pancreas-centered patches cropped from CT images and automatically predict pancreas status.Model performance was evaluated using the receiver operating characteristic（ROC）analysis.The diagnostic performance of DL model was also compared with that of a junior and a senior radiologist in the dataset including atypical cases（106 AP and 106 NAP）.Class-specific discriminative regions were visualized with the heatmap using a Grad-CAM method.Chapter 2:A total of 978 AP patients from Department of Integrated Traditional Chinese and Western Medicine,West China Hospital of Sichuan University from July 2016 to August 2019 were retrospectively recruited,and the severity of AP was determined according to the Revised Atlanta Classification.The NECT images within 72 hours of onset were collected,and clinical factors selected by multivariate logistic regression were used to develop the clinical model via decision tree method.Two Mobile Net V2-backbone based DL models were constructed by using CT images（image model）or both CT images and clinical factors（combined model）,respectively.All models were trained and fine-tuned in the development cohort（n=783）and their performance were evaluated in the independent testing dataset（n=195）.Confusion matrix was used to evaluate the efficacy of the models in the classification of AP severity,and ROC curve was used to analyze the classification efficacy of the three models in the diagnosis of mild AP（MAP）vs others and severe AP（SAP）vs others.Chapter 3:Two hundred and seven AP patients diagnosed with PI after necrosectomy or percutaneous catheter drainage in our hospital from October 2016to July 2019 were retrospectively collected,and 207 cases of non-PI AP patients admitted to our hospital during the same period were matched one by one according to the age,gender and disease severity of the above PI patients.Abdominal ontrast-enhanced CT images and clinical information before these patients were diagnosed with PI were collected.The clinical model was constructed according to the clinical factors screened by univariate and multivariate logistic regression,and the image model based on the Mobile Net V2 architecture was built on the CT images.Subsequently,the combined model was developed by combining CT images and clinical factors.The performance of these models was evaluated in terms of discrimination and clinical usefulness.In addition,the diagnostic efficacy of the combined model was also compared with several classical clinical and imaging scoring systems.Results:Chapter 1:The DL model had achieved a sensitivity,specificity,and AUC of80.82%,81.34%,and 0.871（95%CI,0.850-0.890）in the training dataset,respectively.The sensitivity,specificity,and AUC of the DL model were 74.04%,78.08%,and 0.810（95%CI,0.759-0.854）in the independent testing dataset,respectively.Decision curve analysis demonstrated the clinical usefulness of the DL model.The DL model showed non-inferior performance to the senior radiologist（p=0.625）and outperformed the junior radiologist（p<0.001）with a very fast diagnostic speed（0.15 sec per-case）.The performance of the junior and senior radiologists improved after the assistance of DL model with accuracy increased from62.23%and 79.25%to 94.81%and 97.22%,while the diagnostic time per-case was reduced from 14.34 sec and 8.92 sec to 4.43 sec and 3.17 sec,respectively.Chapter 2:The results of univariate analysis and multivariate logistic regression showed that 24h SIRS、48h SIRS、WBC、NLR、CRP、Ca²⁺、LDH and Cr were the clinical factors significantly related to the severity of AP（P<0.05）.In the validation set,all the three models could predict the severity of AP to a certain extent.Among the three categories of prediction,the accuracy of clinical model,imaging DL model and combined model in predicting SAP was 29.0%,64.5%and 74.2%,respectively;the accuracy of predicting moderate severe acute pancreatitis（MSAP）was 43.2%,27.0%and 32.4%,respectively;and the accuracy of predicting MAP were 80.0%,76.7%and 90.0%,respectively.The overall prediction accuracy of the combined model was the highest among the three models（65.6%）,and the clinical model and imaging DL model were 57.9%and 55.9%,respectively.In the dichotomy,the AUC of the combined model is slightly higher than that of the other two models,whether it predicts MAP vs.others or SAP vs.others.When predicting SAP vs.others,the sensitivity of the three models were all good,among which the clinical model and imaging DL model had the highest sensitivity,both of which were 93.55%,and the combined model was 90.32%.The combined model had the highest specificity（82.93%）,followed by the imaging DL model（76.83%）and the clinical model（67.68%）.The clinical model had the highest specificity（80.00%）and lowest sensitivity（60.95%）in predicting MAP vs others.The sensitivity of combined model was the highest（84.76%）,but the specificity was low（66.67%）.Both the sensitivity and specificity of image DL model were moderate,which were 79.05%and 64.44%,respectively.Chapter 3:According to univariate analysis and multivariate logistic regression,pancreatic necrosis,bubble sign,highest SOFA,days of parenteral nutrition and CRP were independent risk factors for PI.In the validation set,the AUC of the clinical model was 0.819（95%CI,0.718 to 0.895）,with sensitivity and specificity of81.40%and 74.36%,respectively.The AUC of the imaging DL model was 0.806（95%CI,0.704 to 0.885）,with sensitivity and specificity of 81.40%and 66.67%,respectively.The AUC of the combined model was 0.852（95%CI,0.704 to 0.885）,and the sensitivity and specificity were 72.09%and 92.31%,respectively.In addition,compared with MCTSI,SIRS,SOFA and Marshall scoring systems,the diagnostic efficacy of combined model was the best among all scoring systems,but there was no statistical difference in the diagnostic efficacy between combined model and the highest SOFA score and the highest Marshall score.Conclusion:This study preliminary confirms that the DL model based on the Mobile Net V2architecture is feasible for AP diagnosis,early severity grading prediction and secondary PI diagnosis.The AP diagnostic model is an efficient auxiliary tool,which may have application value in emergency medical units,areas with insufficient medical conditions and assisting doctors with low years of experience to make rapid and intelligentized diagnosis of AP.The AP severity grading prediction model can be applied in the early stage of disease,and the combined model which integrated imaging and clinical information achieved a better predictive performance,providing a basis of early risk stratification of the AP patients.The PI diagnostic model is expected to be a noninvasive biological indicator for clinical application.