A deep learning model, designed to predict a patient's conversion from mild cognitive impairment (MCI) to dementia, may offer a convenient and efficient clinical solution for early diagnosis of the devastating illness.
Hong
Patients with MCI have a 10% to 15% risk of converting to Alzheimer's disease (AD) making early identification of MCI important for planning adequate treatment, according to Eun Kyoung (Amy) Hong, MD, radiologist at VUNO, Inc. and PhD candidate at Seoul National University Hospital in South Korea and University of Maastricht in the Netherlands.
"There have been several neuroimaging biomarkers introduced to predict conversion from MCI to dementia, however most of them are expensive, invasive and not readily available in clinical practice," Dr. Hong said.
As an alternative to more expensive modalities, Dr. Hong noted that structural MRI (sMRI) features like medial temporal lobe and hippocampus atrophy are useful in predicting conversion from MCI to dementia, but their evaluation requires a high level of expertise and is prone to high inter-reader variability.
She and her colleagues sought to develop an accurate and reproducible approach using computer-aided risk classification based on structural imaging of the brain.
Comparable Results Between sMRI and Amyloid PET
The team performed a retrospective study of 284 MCI patients who underwent 3D T1-weighted MRI. Of those included, 144 had early MCI (EMCI) and 140 had late MCI (LMCI). From this population, nearly 20% of patients converted from MCI to AD within three years of MCI diagnosis.
The researchers used a previously developed, deep learning-based AD prediction model trained to analyze structural MRI using scans from a total of 1,100 subjects: 550 clinically diagnosed dementia patients and 550 cognitively normal patients.
Training was performed using a deep neural network architecture for computer vision, a field of artificial intelligence (AI) that enables computers to "see" by gathering meaningful information from digital images and other visual input.
Jinyoung Kim, Eunpyeong Hong, Jinkyeong Sung, Wooseok Jung, Changhyun Park, Hyunwoo Oh, Dongsoo Lee, Eun Kyoung Hong from VUNO Inc.
The output from the model was presented as an AD probability score, and scores from the deep learning model were compared between the EMCI and LMCI patients and between patients who converted to AD versus those who did not.
Then, the team analyzed the performance of the AD probability score in predicting conversion of MCI to dementia and compared it with that of brain imaging using amyloid PET which detects the build-up of beta-amyloid plaques in the brain, widely thought to play a central role in AD.
The resulting accuracies for the AD probability score and amyloid PET in predicting conversion from MCI to dementia were comparable, and no statistically significant difference in prediction performance was noted between the two modalities.
"It was interesting to see how well the AD probability score differentiated converters from non-converters in EMCI, LMCI and all MCI patients included in the study," Dr. Hong said, "Moreover, the performance of the AD probability score was comparable to amyloid-beta PET, known to be one of the most sensitive biomarkers for dementia."
Looking ahead, Dr. Hong and her team will continue to train, validate and optimize the model for improved performance in clinical use for early diagnosis and treatment of dementia, as well as to expand its development and use the algorithm for other types of dementia in the near future.
The poster presentation, "Introduction Of MRI-based AI Model In Prediction Of MCI Conversion To Dementia: Could It Be A Key To Early Diagnosis Of Alzheimer's Disease?," (NR03-C8) will take place on Tuesday, Nov. 29 at 4 p.m. Learn more at Meeting.RSNA.org.
