Page 89 - Contributed Paper Session (CPS) - Volume 4

CPS2131 Philip L.H. Yu et al.
patient without any side of weakness was not considered in model
construction. Total Among 120 training subjects, 95 patients (79.16%) were
recorded as side of weakness, and 3 patients suffer both sides of weakness.
Thus, total 588 slices were fed into model training, and only 84 slices
containing MCA sign ground truth.

Side of Weakness w/o Side of
Weakness
MCA 63 0 63
w/o MCA 32 25 57
95 25 110
Table 1. MCA vs Side of Weakness in Training Data

Our model achieves dice similarity coefficient 0.686 on the testing data.
The result is satisfied since the MCA sign is extremely small and quite hard to
gain relative high DSC.

5. Discussion & Conclusion
MCA dot signs are extreme small in size and quite low signal to noise ratio,
the essential step in MCA segmentation task is the localization of specified
regions of interest, which would largely ignore that irrelevant information.
The DSC is not high enough mainly from two aspects. The first is due to its
size is pretty small that even a subtle miss would cause large effect on the
prediction. Total positive ground truth label is less than 0.1% of negative label.
Another is the false positive predictions accounts for large inaccuracy that
largely due to proximity of bone and the similarity to normal age-related
vascular calcification. However, our prediction has high sensitivity that able to
right predict all MCA dot signs in our testing case.
Further post-processing step to distinguish MCA dot signs and false
positive predictions will largely help enhance the model’s performance.
Adequate data with more positive labelled ground truth will enable model to
learn more and become more robust, as well.
Overall, we present an automated method for identifying the hyperdense
MCA dot sign on Noncontrast CT scans. The study can be further reinforced
with additional data input.

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