CP-HRIC · Model · cp-hric-mil family

CP-HRIC-MIL · multi-task model

Shared per-cell layers with one head per question (genetic subtype, APL), trained together so both answers come from one model. The genetic subtype is scored on the Munich folds (every patient tested once); APL on DeepAPL's 24 held-out patients, never used for training or tuning. For MILLIE's test, the same recipe is re-run on MILLIE's own 3-fold split of all 106 Johns Hopkins patients.

Cell encoder
our own encoder (v3e), its last two blocks fine-tuned together with the heads; the rest frozen
Trained on
the Munich patients plus DeepAPL's 82 Johns Hopkins training patients
The Layer 8 robot pointing at a glowing DNA helix beside a tablet

Against the three papers

Each paper’s own test, measured its way.

  1. SCEMILAmacro-F1 over the five subtypes, the paper’s five patient-level folds · paper 0.810.841± 0.019 · 10 seeds+0.031 vs 0.81Beats
  2. MILLIEAUROC 3-fold cross-validation · paper 0.96 ± 0.020.926± 0.031 · 10 seeds−0.034 vs 0.96Matched
  3. DeepAPLAUROC on 24 held-out patients · paper 0.86 (published score; model not re-run)0.869± 0.034 · 10 seeds+0.009 vs 0.86Matched within margin of error

How it got here

Key changes.

  • Step 9 frozen encoder (j6) → Step 10 Same data, encoder fine-tuned with the heads (f1)SCEMILA 0.836 → 0.841 · MILLIE 0.906 (5 seeds) → 0.926 · DeepAPL 0.815 → 0.869
Show all steps

Each step is a change to the model or its data, in order. Bar = mean across seeds, whisker = ± the spread across seeds, diamond = the seeds’ models combined, dashed line = the paper. The axis starts at 0.5.

SCEMILA · Munich · genetic subtype · macro-F1 SCEMILA (0.81)
  1. Step 9 Munich + Johns Hopkins, one model for both questions · frozen encoder (j6)10 seeds · ± 0.028Beats0.836+0.026
  2. Step 10 Same data, encoder fine-tuned with the heads (f1)10 seeds · ± 0.019Beats0.841+0.031
MILLIE · Johns Hopkins · APL · 3-fold CV · AUROC MILLIE (0.96)
  1. Step 9 Munich + Johns Hopkins, one model for both questions · frozen encoder (j6)5 seeds · ± 0.025 · 5 models combined: 0.938In progress0.906−0.054
  2. Step 10 Same data, encoder fine-tuned with the heads (f1)10 seeds · ± 0.031 · 10 models combined: 0.967Matched0.926−0.034
DeepAPL · Johns Hopkins · APL · 24 held-out · AUROC DeepAPL (0.86)
  1. Step 9 Munich + Johns Hopkins, one model for both questions · frozen encoder (j6)10 seeds · ± 0.036 · 10 models combined: 0.826Approaching0.815−0.045
  2. Step 10 Same data, encoder fine-tuned with the heads (f1)10 seeds · ± 0.034 · 10 models combined: 0.875Matched within margin of error0.869+0.009

How much it learned from: the training data, image counts and sizes →

Beats = above the paper by more than run-to-run noise; Matched within margin of error = above it by less than that; Matched = our mean ± spread overlaps the paper’s range; Approaching = within 0.05; In progress = further away. The chip takes the model’s family colour. Every number is the declared configuration's mean across seeds ± the spread across seeds; a combined model (several seeds' models combined into one) is shown after it, never instead of it. Johns Hopkins scores use the cleaned data (prep v2). "Not run · queued" cells are scheduled to be filled.