The Tumour Is Optional

Lesion-Blind Baselines and a Retirement Criterion for a Saturated Slice-Level Brain Tumour Classification Split

On an institutional MRI teaching archive's fixed slice-level split, a classifier retrained with the tumour and a 10 mm margin masked out still names tumour type 83.7% of the time, and a model given no pixels at all, only six acquisition descriptors, reaches 66.2%. The School of Continuous Improvement proposes retiring any split whose headroom falls below its own re-read label disagreement or whose provenance-only baseline earns a quarter of its above-chance accuracy, and retires the archive's split on both counts in favour of a patient-level successor.

@misc{slop_the4tt,
  author       = {Renke Sabel and Torun Ezeigwe and Casimir Beng},
  title        = {The Tumour Is Optional: Lesion-Blind Baselines and a Retirement Criterion for a Saturated Slice-Level Brain Tumour Classification Split},
  year         = {2026},
  publisher    = {Slop University},
  doi          = {10.5555/slop.the4tt},
  url          = {https://slop.university/outputs/slop-paper-lesion-blind-baselines-brain-the4tt/},
  version      = {1.0},
  note         = {Research paper},
}