covtoken / paper /venue_notes.md
Chucks90's picture
covtoken: label-free lesion-subspace token economy (reframed) + gated eval + paper draft
3510f1d verified
|
Raw
History Blame Contribute Delete
2.11 kB

Venue targeting

Option A — Full method paper (recommended primary)

Venue: MICCAI / IEEE TMI / Medical Image Analysis. Framing: label-free lesion subspaces for token-economical medical imaging; the negative result is §6, not the headline. Sells the cross-modality/backbone generality + certificate + routing. Why it survives review: every claim is gated with locked thresholds and CI tests; the precondition sentence (helps where density localizes the lesion; liver = characterized failure) preempts the obvious probe; the negative result demonstrates rigor rather than weakness.

Option B — Negative-results / short paper (strong standalone)

Venue: MIDL (short / negative results track), or a "lessons learned" workshop, or NeurIPS ML4H. Title: "Rank-Based Coverage Objectives Fail for Rare-Lesion Retention: A Mechanism." Core: NEGATIVE_RESULT.md — the ablation + the spanning-vs-concentration mechanism + three convergent lines (ablation, Gate-2 tie, flat C*). Transferable warning for RankMe-flavored medical SSL. A negative result with a mechanism gets cited; this is that.

Option C — Mid-layer probe note

Venue: SPIE Medical Imaging (aligns with the existing representation-coverage probe paper). Core: Finding 1 (lesion signal lives mid-layer; final 0.565 → block-3 0.871). Reconcile with the existing SPIE probe paper: if that probe read final-layer features it was looking in the wrong place; merge or cross-cite so the two reinforce and do not read as salami-slicing. The covtoken negative result gives clean topical separation between the two.

Recommended split

A as the main paper; fold Finding 1 into the SPIE probe paper (C) with cross-citation; keep B in reserve if A's reviewers want the negative spun out. Do not publish A and B as overlapping positives — B is the negative slice of A, clearly fenced.

Figures (in figures/)

  • fig1 layer ablation (Finding 1) - fig2 cross-modality AUROC (Finding 2)
  • fig3 pruning gains (Finding 3) - fig4 floor ablation / negative (Finding 4)
  • fig5 conformal retention certificate