Machine learning software can easily gauge brain atrophy

Artificial intelligence can quickly and accurately rate the amount of atrophy in parts of the brain, which can help in the diagnosis and research of dementia.

A new study reported in arXiv.org, part of the Cornell University Library, is suggesting that the assessment of structural changes in the brain can be made clinically by visually rating brain atrophy from radiological images. It is an inexpensive way to quantify brain atrophy and helps improve the specificity and sensitivity of diagnosis.

These visual atrophy ratings aren’t used widely because the ratings are inherently subjective, researchers note. Radiologists performing them need to be experienced in such work, and the assessments are tedious and time consuming—it takes several minutes per image for a neuroradiologist to perform the rating. And while the rating is potentially useful in a clinical setting, the current process does not easily enable the study of larger groups of images, which is important for research purposes.



The study authors, from Karolinska University Hospital in Stockholm and elsewhere, developed an algorithm based on convolutional and recurrent neural networks to automatically predict the brain atrophy ratings in regions of the brain that are often affected in dementia, such as the hippocampus. Their goal was to provide faster, more objective and more reliable predictions of established visual rating scales of atrophy.

They trained their software model, called Automatic Visual Ratings of Atrophy (AVRA), on 2,350 visual ratings made by an experienced neuroradiologist. They also trained the algorithm to rate MRI images with as little preprocessing as possible to mimic a clinical situation.

Also See: Telemedicine provides second opinions to brain tumor patients

Prior automatic or semi-automatic methods to quantify brain atrophy can’t necessarily measure volume or shape; those that only measure volume can only be used with certain kinds of MRI images, according to the researchers.

The AVRA model’s rating agreements to the radiologist’s ratings were “substantial” and more in agreement than other automatic rating models.

The algorithm also was faster and more reliable, making it more adaptable to real-world use.

“This tool runs in under one minute on a regular laptop, which enables automatically rating thousands of images in a couple of hours. Rating an MRI image of the brain requires minimum amount of preprocessing and the models were built to potentially work in a clinical setting. The main advantage of an automatic model is the absence of randomness, which can ensure rating consistency between different clinics, research groups and cohorts. Thus, AVRA has potential to function as a clinical aid, and to increase the use of visual ratings in research,” the study authors stated.

Moreover, this algorithm may be more versatile than other applicable software. Since the AVRA model predicts ratings directly from voxel intensity values from the images, not just volume measures, the algorithm should work not only with thin MRI images but also thicker MRIs and even the cheaper and more commonly used CT images, from which volumes generally can’t be computed, said the study authors.

The researchers plan to offer the software to neuroradiologists and neuroscientists. “We conclude that automatic visual ratings of atrophy can potentially have great clinical and scientific value, and aim to present AVRA as a freely available toolbox,” they stated.