Ensembled Deep Neural Network for Intracranial Hemorrhage Detection and Subtype Classification on Noncontrast CT Images

Abstract

Rapid and accurate diagnosis of intracranial hemorrhage is clinically significant to ensure timely treatment. In this study, we developedanensembleddeepneuralnetworkforthedetectionandsubtypeclassificationofintracranialhemorrhage.Themodelconsistedoftwoparallelnetworkpathways,oneusingthreedifferentwindowlevel/widthsettingstoenhancetheimagecon-trastofbrain,blood,andsofttissue.Theotherextractedspatialinformationofadjacentimageslicestothetargetslice.BothpathwaysexploitedtheEfficientNet-B0asthebasicarchitectureandwereensembledtogeneratethefinalprediction.Classacti-vationmappingwasappliedinbothpathwaystohighlighttheregionsofdetectedhemorrhageandtheassociatedsubtypes.ThemodelwastrainedandtestedusingIntracranialHemorrhageDetectionChallenge(IHDC)datasetlaunchedbytheRadiologicalSocietyofNorthAmerica(RSNA)in2019,whichcontained674,258headnoncontrastscomputertomographyimagesacquiredfrom19,530patients.Anindependentdataset(CQ500)acquiredfromanotherinstitutionwasusedtotestthegeneralizabilityofthetrainedmodel.Theoverallaccuracy,sensitivity,andF1scoreforintracranialhemorrhagedetectionwere95.7%,85.9%,and86.7%onIHDCtestingdatasetand92.4%,92.6%,and93.4%onexternalCQ500testingdataset.Theheatmapsbyclassacti-vationmappingsuccessfullydemonstrateddiscriminativefeatureregionsofthepredictedhemorrhagelocationsandsubtypes,providingvisualguidanceforradiologiststoassistinrapiddiagnosisofintracranialhemorrhage.