CP-HRIC · Experiments

Every experiment, one page each.

Each experiment asks one question of one public dataset and checks the answer against a published paper, measured the paper’s own way. Open any card for the full experiment: the data, the protocol, and every model’s result.

Two Layer 8 robots in a lab: a green one at a microscope pointing at a DNA display, an orange one reading a tablet
  • Experiment 001CP-HRIC-MIL: BeatsCP-HRIC-S1: Matched

    Which genetic subtype?

    Five answers: PML::RARA, NPM1, CBFB::MYH11, RUNX1::RUNX1T1 or healthy control.

    Data
    TCIA AML Cytomorphology (Munich): 189 patients, 81,214 blood cells. The paper’s own five patient-level folds.
    Paper
    SCEMILA, Hehr et al. 2023: macro-F1 0.81
    • CP-HRIC-MIL · multi-task modelMunich · 189 patients · SCEMILA's folds · macro-F10.841 ± 0.019 · 10 seeds +0.031 vs SCEMILABeats
    • CP-HRIC-S1 · multi-task modelMunich · 189 patients · SCEMILA's folds · macro-F10.809 ± 0.016 · 10 seeds −0.001 vs SCEMILAMatched
    Open the full experiment
  • Experiment 002CP-HRIC-MIL: MatchedCP-HRIC-S1: In progress

    APL or another AML?

    One question, at a second hospital: acute promyelocytic leukaemia (APL, the PML::RARA answer above) or any other AML.

    Data
    Johns Hopkins deepAPL: 106 patients (34 APL, 72 other AML). Two published ways to test it: random 3-fold cross-validation, and DeepAPL’s fixed split of 82 training patients and 24 held-out patients.
    Paper
    MILLIE, Manescu et al. 2023: AUROC 0.96 ± 0.02 by 3-fold cross-validation · DeepAPL, Sidhom et al. 2021: AUROC 0.86 on the 24 held-out patients
    • CP-HRIC-MIL · multi-task modelMILLIE's test · 3-fold CV · 106 patients · AUROC0.926 ± 0.031 · 10 seeds −0.034 vs MILLIE · 10 models combined: 0.967 · shortcut-only floor 0.634 ± 0.031 for this protocolMatched
    • CP-HRIC-MIL · multi-task modelDeepAPL's test · 24 held-out · AUROC0.869 ± 0.034 · 10 seeds +0.009 vs DeepAPL · 10 models combined: 0.875 · shortcut-only floor 0.608 for this protocolMatched within margin of error
    • CP-HRIC-S1 · multi-task modelMILLIE's test · 3-fold CV · 106 patients · AUROC0.885 ± 0.025 · 10 seeds −0.075 vs MILLIE · 10 models combined: 0.944 · shortcut-only floor 0.634 ± 0.031 for this protocolIn progress
    • CP-HRIC-S1 · multi-task modelDeepAPL's test · 24 held-out · AUROC0.864 ± 0.038 · 10 seeds +0.004 vs DeepAPL · 10 models combined: 0.917 · shortcut-only floor 0.608 for this protocolMatched within margin of error
    Open the full experiment

Numbers are means across independently trained seeds (± spread) unless marked “1 run”; “combined” means those seeds’ models combined into one. Genetic subtype is scored in macro-F1 and APL in AUROC (1.0 is perfect; chance is 0.20 and 0.50). Deltas are our figure minus the paper’s reported figure on the paper’s own patients. We compare against DeepAPL’s published number; we did not re-run DeepAPL’s own model.

Experiment 004One cell-type tree for every dataset. The first step past the papers: one head names the cell type for all five public datasets. Kept.

Model data

How much our models learned from.

  • 52.1 GBimage intake as decoded pixels · 316,415 images: the encoder's images plus the Munich and Johns Hopkins patients' cells, including the 24 held-out Johns Hopkins test patients
  • 51.9 GBtraining only, as decoded pixels · 313,420 images: the encoder's images plus the Munich patients' cells and the 82 Johns Hopkins training patients' cells; the 24 held-out test patients excluded
  • 23.2 GBthe same images as image files on disk

What they learned from

  • MLL23 peripheral blood cellstrains our cell encoder41,621 images10.4 GB decoded
  • Bone Marrow Cytomorphology, Munichtrains our cell encoder (17 of 21 cell classes so far; damaged files excluded from training)122,313 images22.9 GB decoded
  • Barcelona peripheral blood cellstrains our cell encoder17,092 images6.70 GB decoded
  • Ostrava white blood cellstrains our cell encoder16,027 images3.99 GB decoded
  • LeukemiaAttri (ITU)trains our cell encoder23,907 images2.18 GB decoded
  • AML-Cytomorphology, Munich (TCIA)Munich patients: trains and tests the models on SCEMILA's folds81,214 images5.05 GB decoded
  • Johns Hopkins DeepAPLJohns Hopkins patients: 82 train, 24 held out for DeepAPL's test14,241 images886 MB decoded

At the encoder’s 144 px input, one training pass over its images is 13.7 GB; a 12-pass run processes 165 GB. The encoder file is 85.3 MB.

Everything on our storage: 4.32 TB · 639,696 files · 24 datasets
  • Pediatric leukemia bone-marrow slides (IDC)downloaded, not yet used2.36 TB2,510 files
  • CPTAC-AML slides (TCIA)downloaded; file-level verified; not yet used1.09 TB272 files
  • Haemorasis blood smears (MDS, anaemia)downloading382 GB192 files
  • CMB-AML (TCIA)downloaded, not yet used265 GB67 files
  • Penn bone-marrow niches (CODEX)downloaded, not yet used67.3 GB31 files
  • Ostrava white blood cellstrains our cell encoder46.2 GB2 files
  • cAItomorph test set (409 patients)downloaded, not yet used21.3 GB201,595 files
  • MD Anderson AML (COMET)downloaded, not yet used12.5 GB25 files
  • AML-Cytomorphology LMU (TCIA)downloaded, not yet used11.8 GB18,368 files
  • C-NMC 2019 (TCIA)downloaded, not yet used10.4 GB15,121 files
  • LeukemiaAttri (ITU)trains our cell encoder9.88 GB25,825 files
  • MLL23 peripheral blood cellstrains our cell encoder9.73 GB19 files
  • Bone Marrow Cytomorphology, Munich (fresh copy)downloaded, not yet used6.96 GB171,375 files
  • Dresden bone marrow (Eckardt)downloaded, not yet used5.29 GB7 files
  • AML-Cytomorphology, Munich (TCIA)Munich patients: trains and tests the models on SCEMILA's folds5.07 GB81,216 files
  • Bone Marrow Cytomorphology, Munichtrains our cell encoder (17 of 21 cell classes so far; damaged files excluded from training)5.00 GB122,325 files
  • WBCBench 2026 (ISBI challenge)downloaded, not yet used3.94 GB23 files
  • UHN AML imaging mass cytometrydownloaded, not yet used3.10 GB2 files
  • MDS marrow multiplexed imagingdownloaded, not yet used1.66 GB2 files
  • DinoBloom hemato embeddings (MarrLab)downloaded, not yet used949 MB707 files
  • Johns Hopkins DeepAPLJohns Hopkins patients: 82 train, 24 held out for DeepAPL's test539 MB3 files
  • Barcelona peripheral blood cellstrains our cell encoder281 MB2 files
  • WBCAtt+ white-cell attributesdownloaded, not yet used190 MB5 files
  • Taleqani ALLdownloaded, not yet used116 MB2 files

The whole programme’s public data, not only these models’. Sizes are as stored: many datasets arrive as archives, so a few files can hold many images. Datasets still downloading are shown at their current size.