Purpose or Case Report: Hypoxic Ischemic Encephalopathy (HIE) is a brain injury caused by a lack of blood and oxygen supply to the brain. HIE affects 4,000,000 term-born neonates per year worldwide, with an estimated 2 billion/year cost in the US, let alone family burdens. Therefore, reducing mortality and morbidity for HIE patients remains an important public health concern. Therapeutic hypothermia (TH) was established in 2005 as a standard therapy by cooling patients to 33-34°C in the first six postnatal hours for 72 hours. However, 35–50% of the patients still experience adverse outcomes, defined as death or cognitive Bayley Scales of Infant Development by age two years. Ongoing HIE-related trials worldwide are testing whether new therapies can supplement TH and further reduce adverse outcomes. However, therapeutic innovation is slow and inconclusive, for 1) before therapy, patients at high risk of developing adverse outcomes cannot be identified; 2) after therapy, outcomes cannot be measured until age two years. Besides, public MRI data exists for hundreds of patients with brain tumors, Alzheimer’s Disease, and other diseases, fueling AI’s success in MRI-based diagnosis and prognosis of brain tumor, Alzheimer’s Disease, and other disorders. In contrast, annotated MRIs with linked clinical and bio-marker data do not exist publicly for HIE. Our previous work has collected multi-site HIE MRI data. Therefore, to fill the gap in HIE diagnosis with MRI data, target high-risk patients, increase efficiency, evaluate therapeutic effects early, and expedite therapeutic innovations, in this work, we propose to predict 2-year neurocognitive outcomes in neonates using brain MRIs by deep learning methods.
Methods & Materials: With representative deep learning networks, we extracted the features from brain MRIs we collected to predict 2-year neurocognitive outcomes.
Results: Experimental results demonstrate that the brain MRIs can help predict neurocognitive outcomes.
Conclusions: Brain MRIs can help predict neurocognitive outcomes. With the encouraging results, next step, we will explore combining other information such as lesion detection to improve the neurocognitive outcome prediction performance.