Efficiency Of A Deep Learning-Based Computer-Assisted Diagnostics System In Spontaneous Intracerebral Hemorrhage Volume And Midline Shift Measurement

Objective To evaluate the performance and accuracy of a deep learning(DL)-based computer-assisted diagnostics(CAD)system in segmenting spontaneous intracerebral hemorrhage(ICH)volume and brain midline shift.Methods We retrospectively reviewed 105 cases of brain scans of patients with acute spontaneous ICH.Depending on the presence of intraventricular hemorrhage(IVH)extension,patients were divided into two groups: ICH without(n = 56)and with IVH(n =49).96 patients with spontaneous supratentorial ICH were evaluated for midline shift.ICH was segmented and measured using a DL-CAD system as well as manual measurements:computed tomography-based planimetry(CTP)and ABC/2 score.The standard ICH volume were measured by CTP,and ABC/2 score only measured hemorrhage volume in ICH without IVH group.The midline shift uses the radiologist’s analysis results as standard to evaluate the detection performance of the DL-CAD system.statistic analysis were used to analyze the differences in volume,length of processing time,and midline shift among the different measurement approaches.Results The mean deviation values between DL-CAD system and CTP were-0.10 ml for ICH without IVH group,-0.11 ml for ICH with IVH group.The 95% LOA values were-4.38 to 4.18 ml,-7.05 to 6.82 ml,respectively;the differences are both not significant(P <0.05).Strong correlations and agreement were observed between CTP and DL-CAD in two groups(r = 0.99,0.99 P < 0.01;concordance correlation coefficient [CCC] = 0.99 and 0.99).In the ICH without IVH group,The mean deviation values were 1.53 ml for the ABC/2score and CTP,1.63 ml for ABC/2 and DL-CAD system score.The 95% LOA values were-7.90 to 10.96 ml and-7.96 to 11.22 ml,respectively,the differences were statistically significant(P < 0.05).The correlation and agreement between ABC/2 score and CTP or DL-CAD system in ICH without IVH group were good(r = 0.99 and 0.97,P < 0.01;CCC= 0.99 and 0.97;respectively).The DL-CAD segmentation took a significantly shorter amount of time than CTP in two groups(P < 0.01),but it was a little slower than ABC/2score in ICH without IVH group(median time difference = 0.18[0.11 to 0.26];P<0.01).The overall deviation of ICH volume measured by ABC/2 score was greater than the DL-CAD system.The sensitivity,specificity and accuracy of the DL-CAD system in predicting brain midline shifts were 67.24%,69.44%,and 68.09%,respectively.The deviation of midline shift measured by DL-CAD system was-0.70(-3.85,0.00)mm,and the difference was statistically significant(P < 0.01).The midline shift measured by DL-CAD system had weak correlation and consistency with the results of manual measurement by radiologists(r = 0.56,P < 0.01;CCC = 0.62).Conclusion We found that the DL-CAD system could measure hematoma volume in patients with acute spontaneous ICH as accurately as and more efficiently than manual measurements.We suggest this as a promising tool to help physicians achieve precise ICH quantification in practice.However,the accuracy of the DL-CAD system for detecting brain midline shifts is still low,and further optimization of the algorithm model and the addition of high-quality training data are needed to improve the accuracy of detection.