Useful slices are sparse
Small abnormalities may be visible only briefly, while the full MRI study contains a large amount of normal or redundant anatomy.
Medical imaging / study classification
MRNet MRI classification.
An exploratory attention architecture that preserves local slice evidence before aggregating it into a volume-aware study prediction.
Interactive architecture
Local evidence, then study context.
Each slice gets its own representation before attention decides how evidence should combine across the scan.
Case study
A diagnosis distributed through a scan
Separate two kinds of attention
Patch tokenization and self-attention encode content within slices; inter-slice aggregation then assigns importance across the volume.
Show the experimental path
The public repository contains exploratory notebooks for architecture, data handling, training experiments, and visual inspection.
Architecture detail
From slices to a study decision
Sample slices
Preserve the ordered MRI sequence as the primary unit of analysis.
Tokenize patches
Represent local structures within each slice through attention-ready visual tokens.
Aggregate evidence
Learn which slice representations should carry weight in the study-level summary.
Classify the study
Convert the aggregated volume representation into the target diagnostic output.
Evidence boundary
A notebook experiment, not a diagnostic product.
The repository records architectural exploration and model-development work. It does not claim clinical validation or deployment readiness.