Feature pyramid network for automatic segmentation and semantic feature classification of spontaneous intracerebral hemorrhage hematoma on non-contrast CT images
10.13929/j.issn.1003-3289.2024.10.007
- VernacularTitle:基于特征金字塔网络自动分割平扫CT所示自发性脑出血血肿并判断其语义特征
- Author:
Changfeng FENG
1
,
2
;
Qun LAO
;
Zhongxiang DING
;
Luoyu WANG
;
Tianyu WANG
;
Yuzhen XI
;
Jing HAN
;
Linyang HE
;
Qijun SHEN
Author Information
1. 杭州市儿童医院放射科,浙江 杭州 310005
2. 浙江大学医学院附属杭州市第一人民医院放射科,浙江 杭州 310003
- Keywords:
cerebral hemorrhage;
hematoma;
tomography,X-ray computed;
deep learning
- From:
Chinese Journal of Medical Imaging Technology
2024;40(10):1487-1492
- CountryChina
- Language:Chinese
-
Abstract:
Objective To observe the value of feature pyramid network(FPN)for automatic segmentation and semantic feature classification of spontaneous intracerebral hemorrhage(sICH)hematoma showed on non-contrast CT.Methods Non-contrast CT images of 408 sICH patients in hospital A(training set)and 103 sICH patients in hospital B(validation set)were retrospectively analyzed.Deep learning(DL)segmentation model was constructed based on FPN to segment the hematoma region,and its efficacy was assessed using intersection over union(IoU),Dice similarity coefficient(DSC)and accuracy.Then DL classification model was established to identify the semantic features of sICH hematoma.Receiver operating characteristic curves were drawn,and the area under the curves(AUC)were calculated to evaluate the efficacy of DL classification model for recognizing semantic features of sICH hematoma.Results The IoU,DSC and accuracy of DL segmentation model for 95％sICH hematoma in training set was 0.84±0.07,0.91±0.04 and(88.78±8.04)％,respectively,which was 0.83±0.07,0.91±0.05 and(88.59±7.76)％in validation set,respectively.The AUC of DL classification model for recognizing irregular shape,uneven density,satellite sign,mixed sign and vortex sign of sICH hematoma were 0.946-0.993 and 0.714-0.833 in training set and validation set,respectively.Conclusions FPN could accurately,effectively and automatically segment hematoma of sICH,hence having high efficacy for identifying semantic features of sICH hematoma.
