Dyslexia and isolated spelling deficits quietly undermine schooling in millions of children of normal intelligence. The earlier they are identified, the more effective intervention becomes. The trouble is timing, and the tooling behind it.
Structural MRI can reveal neurobiological signatures of reading difficulty, but training a reliable classifier needs large, well-labeled datasets. Assembling those is extraordinarily hard when subjects are children and the diagnostic categories are subtle.
A study in Neural Computing and Applications works around that bottleneck. Researchers S. Santhiya and C. S. Kanimozhiselvi at Kongu Engineering College in Tamil Nadu built a framework called MedFusionNet that attacks the data problem from two directions at once.
Two fixes for one small dataset
The first is generative. Rather than waiting for thousands of scans, the team synthesized additional training data to expand a small pool. The second is architectural, fusing several neural network designs into one decision-making system so no single model carries the whole judgment.
The dataset, drawn from the public OpenNeuro repository, held just 58 high-resolution scans: 22 typically developing children, 16 with isolated spelling deficits and 20 with dyslexia. Fifty-eight images is a vanishingly small amount for deep learning, and the class imbalance makes it harder still.
Despite that, MedFusionNet classified pediatric brain MRI scans with 96 percent accuracy. The approach suggests a path for other rare or hard-to-collect pediatric datasets, where generative augmentation and ensemble fusion together substitute for volume that simply does not exist.
