Title: Biomedical Informatics Opportunities in Pediatric Neuroimaging
Hosted by: Department of Biomedical Informatics (DBMI), University of Pittsburgh
Speaker: Rafael Ceschin
Summary and thoughts:
Rafael Ceschin from the Children’s Hospital in Pittsburgh presented this talk about his ongoing and future work in pediatric informatics. It was a very informative and interesting lecture, specially for someone who has not worked in or read about the area before. Pediatric informatics in this case focused heavily on neuroimaging informatics i.e. analyzing and classifying MRI scans from neonates. While neuroimage classification is generally a difficult task, it presents unique challenges when applied to pediatrics. The major challenges in the area include data availability and acquisition (as is usually the case) along with the challneges present in the data once it is acquired. Generally, there are 2 ways to get datasets for machine learning/deep learning solutions - existing datasets with both cases and controls, and prospective studies to collect data. Existing datasets with MRI scans of neonates are (understandably) very limited. A common issue is not having appropriate controls in the data as healthy children are not often required to have MRI scans. Prospective data collection has its own challenges wherein you must convince parents/guardians to get MRI scans, gamble with the newborns staying still through a long and loud process of collecting the scan and then dealing with the structural brain changes that occur more frequently in newborns. Dr. Ceschin did not shy away from delineating the many challenges in the field and emphasized that these are what make the problem unique and worth solving.
A major area of work includes classification of scans of cerebellum and hippocampus in the brain to diagnose congenital disorders. A neural network pipeline with custom segmentation and 3-dimensional convolutional neural nets (CNN) (examples include NeBSS and CRBNet projects) have provided moderate success in the task of classifying cerebellum scans, although there is a long way to go. Classification of hippocampus images is much harder due to inherent structural complexities, smaller area, more variation in the structure across subjects, and malrotation. The biggest need for the problem is harmonization of the data once it is acquired. Different variety and brands of scanners introduce bias in the dataset which is hard to decipher and makes the classification more difficult. Thus, before any modeling, the data must be preprocessed or harmonized and all biases identified to not skew the results. Some ways to achieve this include deep learning (DeepHarmony project), transfer learning (scanning adult phantons with different scanners to categorize the biases), and quantitative harmonization. However, when the size of dataset is very small, the usual techniques do not work so well out of the box. Dr. Ceschin has applied the Hierarchical Bayes technique to combine independent datasets and use simulated data to estimate scanner bias. While his work has been halted briefly due to COVID, they continue to develop new techniques and models, as well as plan for future work to integrate the AI models into clinical workflows with the minimum disruption to the clinicians.