MR Spine Sequence Classifier

Spine MRI sequence descriptions are free-text, vendor-dependent and often wrong — and every downstream model depends on knowing T1 from T2.

A ResNet-18 whose first convolution accepts N central slices as channels — a lightweight 2.5-D approach that reuses a pretrained 2-D backbone instead of a full 3-D CNN. Frozen feature extraction was systematically compared against full fine-tuning; a sweep settled on 10 slices. Refactored into a modular, config-driven pipeline.

• Macro accuracy and F1 from 0.80 to 0.92 by unfreezing the backbone; T1 precision 1.00.
• Dataset scaled ~8.5× (600 → 5,100 series) with class and sex balance preserved.
• Became the eligibility gate for the lumbar spine measurement pipeline.

• Required the frozen-vs-fine-tuned comparison be measured, not chosen by convention — the decision behind the 12-point gain.
• Directed reuse as a hard gate in the device pipeline rather than a standalone research model.