Georgia residents who have skin cancer may have heard of convolutional neural networks. These are networks that allow machines to analyze visual imagery, in this case images of possible cancers. Their accuracy may even equal that of a dermatologist, but much remains to be developed.
Adversarial attacks on CNNs
Flaws in the CNNs themselves have led them to misclassify or misdiagnose melanoma lesions in particular. For a while, experts would give the systems melanoma images as well as images of moles (melanocytic naevi), and when they saw where the system would go wrong, they retrained the systems with different data.
This led them to see that two things in particular, changing an image’s color balance and rotating it, contribute to the wrong classifications. These are considered “adversarial attacks” because they deliberately cause misinterpretations by inputting novel data.
Improving the situation
Improving color balance and the way an image is rotated or translated can help improve the performance of CNNs, and experts are currently coming up with other ways to avoid issues. For example, retraining the systems with those adversarial images may help. They also found that the ability of a CNN to diagnose an image varies based on what device the image was taken on.
A lawyer for victims of misdiagnosis
The human eye is still more reliable when it comes to interpreting images, but even here, doctors may fail through some negligent act on their part. Under medical malpractice law, victims of negligence are entitled to some compensation, at least that which covers their medical bills, lost wages and pain and suffering. To see how strong of a case you have, you may want to meet with an attorney and discuss your situation.