A new dimension for deep-learning diagnoses

Accurate, rapid and automated diagnoses of acute neurological events and retinal disease based on 3D medical images can be provided by new deep-learning algorithms, report two separate studies published online this week in Nature Medicine.

Deep-learning approaches can diagnose various diseases from 2D medical images, but whether this could also be achieved on complex, detailed 3D images has remained unclear. As volumetric imaging contributes to many medical diagnoses made by practising clinicians, the successful implementation of deep-learning algorithms to make diagnosesvia 3D images may bring this technology one step closer to achieving performance equivalent to a human expert.

Eric Oermann and colleagues analyze more than 37,200 head computed tomography (CT) scans using a new convolutional neural network approach. They accurately diagnose the presence of acute neurological events, such as strokes or hemorrhages, and show that this system could accelerate diagnosis times by simulating a clinical application.

In a separate study, Olaf Ronneberger and colleagues develop a deep-learning architecture to analyse optical coherence tomography (OCT) eye scans and diagnose retinal disease with 95% accuracy. Their system performs separate 3D image segmentation and disease diagnosis, which enables accurate performance on complex medical scans acquired by different imaging devices.

These complementary studies successfully apply deep-learning algorithms to the rapid analysis of 3D medical images, which suggests that these systems can potentially improve clinical workflows by providing fast but accurate diagnoses.