Cause of the TEA recall peak under gap-penalty ablation in MRI

Investigate the cause of the small peak in thoracic enumeration anomaly recall observed around skip-penalty values in the interval [0.75, 1.00] during the MRI vertebra-gap ablation with random removal of one vertebra, and ascertain whether this behavior reflects a systematic property of the sequence predictor or is attributable to random noise.

Background

In the vertebra-gap ablation, the authors allowed non-consecutive label sequences and introduced a single random gap per subject in the MRI test set. They varied a penalty for predicting gaps and measured performance metrics, including TEA recall.

They observed a small TEA recall peak for gap-penalty values around [0.75, 1.00], but could not explain this behavior, suspecting random noise. Understanding whether this peak is systematic or incidental may inform the sequence predictor’s behavior under missing detections.

References

Interestingly, the TEA rec. has a small peak around the interval [0.75, 1.00] in the MRI sequence, for which we can offer no explanation other than random noise.

— VERIDAH: Solving Enumeration Anomaly Aware Vertebra Labeling across Imaging Sequences  (2601.14066 - Möller et al., 20 Jan 2026) in Section 5 (Ablations), Subsubsection "Vertebra Gaps"