Page 88 - Contributed Paper Session (CPS) - Volume 4

CPS2131 Philip L.H. Yu et al.
Our deep learning model used Adam optimizer with 1e-5 initial learning rate
and trained on 200 epochs with Tesla K80 GPU.
To evaluate the performance of the deep learning architecture in
segmenting the hyperdense MCA dot signs, the dice similarity coefficients
(DSC) is utilized as the evaluation metric for goodness of fit. The DSC is
defined as
2| ∩ |
=
|| + ||
where A and B represent the regions of all voxels of ground truth and
segmentation respectively.

Fig. 4. Deep Learning Architecture

4. Result
Within 150 CT scans, 74 patients were diagnosed to have hyperdense MCA
dot sign and the rest were empty. Among the 74 positive MCA sign subjects,
63 were within training and the rest 11 were in the testing. As shown in Table
1, patients without side of weakness do not have MCA dot sign; thus the model
involved side of weakness and a filter to select the potential subjects might
have MCA. We subtract cropped bounding boxes within specified regions of
interest for each set of CT scans. If patient suffer side of weakness, we extract
his or her ROI within corresponding left or right hemisphere; if patient suffer
both side of weakness, we use both ROIs in two hemispheres; otherwise,

77 | I S I W S C 2 0 1 9