Fig 1.
Examples of nCCT images, (a)-(d), and the corresponding CTp, (e)-(h).
The color shows the cerebral blood volume (CBV), cerebral blood flow (CBF), the mean transit time (MTT) generated by CTp, the ischemic area- blue, the dead tissue-red.
Fig 2.
Three basic steps of the proposed model 1) label assignment, 2) feature map extraction, 3) infarct localization.
Fig 3.
CTp with the ground truth, ncCT aligned with the CTp.
Fig 4.
The architecture of the input for the proposed DNN.
Fig 5.
Input for feature extraction (a) brain region on an ncCT slice (b) corresponding CTp, (c) the mask of the brain region.
(a) ncCT Slice. (b) Label. (c) Mask.
Fig 6.
CTp and ncCT slices, and the result of aligning CTp with ncCT.
Fig 7.
(a) DeepLabv3+ with. (b) DeepLabv3+ with MobileNet ResNet50. (c) DeepLabv3+ with. (d) Pixel-wised Feature ResNet101 and Neural Network.
Fig 8.
Localizing the infarct region using the extracted feature maps.
Table 1.
Performance of the model vs. the size of the block and post processing.
Fig 9.
Precision and recall values of Ensemble3 (left) and Ensemble5 (right) with different classification threshold.
Table 2.
Testing the feature maps and input blocks.
Fig 10.
Accuracy (a), precision (b), recall (c), and F1 score (d) of Ensemble model with 3- and 5- blocks by MobileNet, ResNet101, and Resnet50.
(a) Accuracy of all model with (left) and without (right) post processing. (b) Precision of all model with (left) and without (right) post processing. (c) Recall of all model with (left) and without (right) post processing. (d) F1 score of all model with (left) and without (right) post processing.
Table 3.
The average IoU of feature maps and input blocks.
Fig 11.
Average IoU of regions obtained from each model using different classification threshold values with postproceessing (left) and without it (right).
Fig 12.
Box plots of IoU for MobileNet, hand-crafted feature map, ResNet101 and Resnet50.
Ensemble model with 1, 3 and 5 extended output blocks. Each plot shows the median values. IoU of all model with (left) and without (right) post processing.
Fig 13.
Lesion detection of the original ncCT (a) label on CTp (b) of the case that achieved the best segmentation result. (c) Ensemble3 (e) Ensemble5 without postproceessing (d) Ensemble3 (f) Ensemble5 with postproceessing.
Fig 14.
Lesion detection (a) CTp, (b), (d) Ensemble3 and Ensemble5 without postproceessing, (c), (e) Ensemble3 and Ensemble5 with postprocessing.
Fig 15.
Lesion detection (a) ncCT and its corresponding CTp (b) of the case that achieved the lowest result due to the failed postprocessing, (b), (d) Ensemble3 and Ensemble5 without postprocessing (c), (e)Ensemble3 and Ensemble5 with postproceessing.
Fig 16.
Lesion detection of the label on CTp (a) that achieve the low result. The result by Ensemble3 and 5 without postproceessing is depicted in (b) and (d), respectively. The result with postproceessing is in (c) and (e) respectively.