Study On Auxiliary Diagnostic Model For Invasive Pulmonary Aspergillosis Based On Machine Learning

The incidence of invasive pulmonary aspergillosis(IPA)has dramatically increased during the past two decades,with an annually increased morbidity and mortality.The gold standard for IPA diagnosis is exclusively pathological examination,however,only few patients were diagnosed pre-mortem owing to the difficulty of invasive specimen sampling.Currently,diagnosis for IPA is mainly dependent on a comprehensive analysis of risk factors,clinical manifestations and chest radiological features of patient with suspected fungal pneumonia,along with a compromised strategy of graded diagnosis and treatment.Although this graded strategy is applicable in clinical practice,it is highly dependent on clinical experience of the attending physicians or radiologists,resulting in a high rate of misdiagnosis.To date,a great deal of efforts has been taken domestically and globally to promote the diagnostic accuracy for IPA,and some innovated diagnostic methods have been introduced to the clinical practice.However,the reliability and accuracy of these methods are still to be approved.We constructed an end-to-end deep learning model(IPA-NET)for early automatic diagnosis of IPA using the computed tomography(CT)images and patient’s clinical characteristics.First,A total 300,000 chest CT images from the website of China National Center for Bioinformation were employed for model transfer learning.Second,chest CT images and clinical information of 74 cases with IPA and 74 matched patients with common pneumonia from Nanfang and Zhujiang Hospital of Southern Medical University were collected and preprocessed.Finally,the preprocessed data were used for model training and verification.The IPA-NET showed a remarkable diagnostic performance for IPA,with an accuracy of 97.4%,the sensitivity and specificity of 0.96 and 0.98,the precision and F1 score of 0.98 and 0.97,respectively.The area under the curve is 0.99(95%CI,0.98 to 0.99).this newly constructed deep learning model provided a noninvasive and reliable method for early diagnosis of IPA.