Convolutional neural networks not diagnosing skin cancer well

Convolutional neural networks are algorithms used in machine learning mainly for the analyzing of visual images. In the medical field, CNNs can analyze images of possible skin cancers. North Carolina residents should know that many experts have touted the effectiveness of CNNs, saying that in their ability to diagnose the cancer, they can approach the accuracy level of a dermatologist. However, there are concerns with this technology.

CNNs misclassifying altered images

In November 2020, a letter to the editor in the Journal of Investigative Dermatology brought to attention certain flaws in CNNs that cause the systems to misclassify and misdiagnose melanoma lesions. The way these neural networks can go wrong is if they are deliberately retrained using completely different images. This is known as an “adversarial attack.”

Other alterations can turn the visuals into potentially confusing novel data. For example, changing the color balance on an image or rotating the image can cause the CNN to misclassify it. Different image-taking devices have different color balance settings and thus influence the accuracy levels.

What experts can do in the future

There are two goals:

1. Retrain CNNs to make them immune to adversarial attacks; and
2. Standardize the way images are taken and presented.

Otherwise, skin cancer patients may continue to be misdiagnosed and suffer harm as a result.

For the victims of a diagnostic error

Cancers are among the most commonly misdiagnosed medical conditions. If you or a loved one suffered physical pain and loss because of a misdiagnosis or delayed diagnosis, there’s a chance that you could file a claim under medical malpractice law. A lawyer may be able to help you prove that a physician or facility was negligent. Investigators may even come in to strengthen your case before all the settlement negotiations begin. To learn more, you may want to schedule a case evaluation.