Featuring Prabhjot Kaur, Ph.D.

Magnetic resonance imaging (MRI) is central to clinical diagnosis for its exceptional soft-tissue contrast, yet its impact is constrained by physics (SNR–resolution trade-offs, artifacts), time (long scans, motion), and access (cost, protocol variability, low-field systems). These constraints can diminish image quality and contribute to missed or uncertain diagnostic findings. This talk traces a methods-to-impact journey across MRI quality, clinical assessment, and pediatric neuroimaging with three aims: (1) improve low-field MRI for rural care; (2) test when deep learning truly surpasses—or should complement—machine learning with MRI-specific priors; and (3) identify, in epilepsy, which acquisition parameters (e.g., voxel size, TR/TE, FLAIR) most influence lesion detectability. 

During my doctoral work, I developed early CNN-based low-field enhancement for super-resolution and contrast translation, incorporating skip connections and physics/structure-aware priors alongside optimized classical ML baselines. After the doctorate, I led automated MRI quality assessment on large, multi-site datasets: CNNs and handcrafted-feature models were compared head-to-head to quantify reliability, define actionable quality thresholds (SNR, CNR, motion), and guide re-acquisition. More recently, my pediatric epilepsy program builds multimodal anomaly detection across sMRI/fMRI/dMRI, augmented by GAN/VAE synthesis to bolster data diversity and weak-label NLP to generate clinician-facing summaries. 

Results indicate that enhanced low-field images increase reader confidence, standardized QA improves acceptance/recall decisions, and current pediatric models produce lesion-concordant explanations with stronger cross-site generalization when trained with synthetic augmentation. I conclude with a pragmatic path that balances ML and DL by data regime; emphasizes deployment pieces—harmonization, interpretability, acquisition guidance, and clinician-ready reporting; and targets a measurable reduction in missed abnormalities while shortening or stabilizing acquisitions—moving neuroimaging AI toward reliable clinical impact.