Conclusion

In conclusion, the neural network algorithm developed in this project has many advantages. First, it performs all image processing and feature extraction automatically, thus making it easy to use. Second, it can achieve significantly high accuracy (97when training and testing on the same CT – with visible stroke) when performing binary classification on CT scans with visible stroke regions. Finally, it produces a clear, concise summary of the classification prediction in the form of an easy to interpret image. However, the algorithm is still severely limited in that it is unable to make accurate predictions for CT scans with non-visible stroke regions. Despite this, the end result of this project not only satisfies the proposed objectives, but also serves as the first step for a future algorithm capable of actually performing well on ambiguous CT scans. There is no doubt that such an algorithm would be of tremendous use to the medical community.