Upload metadata.txt with huggingface_hub

metadata.txt

@@ -0,0 +1,89 @@
+================================================================================
+ Toxoplasma gondii — Parasitophorous Vacuole (PV) Segmentation
+ CellposeSAM (cpsam_v2) fine-tuned model
+================================================================================
+
+WEIGHT FILE      : cpsam_v2_toxo_r2   (1,218,643,463 bytes ~1.16 GB)
+TASK             : 2D instance segmentation of intracellular Toxoplasma
+                   parasitophorous vacuoles / parasite clusters in fluorescence
+                   microscopy.
+ARCHITECTURE     : Cellpose-SAM
+BASE MODEL       : cpsam_v2   (cellpose 4.2.1.1; MODEL_NAMES includes cpsam_v2)
+ROUND            : 2 (production). Round 1 kept separately.
+AUTHOR           : Einar Olafsson, Carruthers Lab
+DATE             : 2026-08-31
+FRAMEWORK        : cellpose >= 4.2 (CellposeModel API)
+
+--------------------------------------------------------------------------------
+HOW TO LOAD
+--------------------------------------------------------------------------------
+  from cellpose import models
+  m = models.CellposeModel(gpu=True, pretrained_model="cpsam_v2_toxo_r2")
+  masks, flows, styles = m.eval(img, normalize=True,
+                                flow_threshold=0.4, cellprob_threshold=0.0)
+
+In spacr: point the CellposeModel / MakeMasks 'pretrained_model' path at this file.
+
+INPUT EXPECTED   : single-channel fluorescence, the red / parasite channel
+                   (e.g. PerkinElmer C02 or a dsRed channel). 16-bit accepted
+                   (the model was trained/evaluated on raw 16-bit with
+                   normalize=True). Typical well crop ~2000x2000 px.
+
+--------------------------------------------------------------------------------
+TRAINING DATA
+--------------------------------------------------------------------------------
+Round 2 total: 229 training images.
+  - 104 NAS gold wells (lab-annotated Toxoplasma vacuole masks)
+  - 125 newly curated wells across additional marker conditions
+        (RH-dsRed, RH-abTg, ME49-abTg, host-cell markers: tubulin / ESCRT /
+         cellmask), curated in the custom Qt tool. Empty wells kept as
+         negatives (min_train_masks=0).
+Held-out (never trained on):
+  - 11 NAS wells  (fixed seed=0 split, identical to round 1 -> comparable)
+  - 40 curated new-condition wells (cross-condition generalisation set)
+Epochs 100, lr 1e-5.
+
+--------------------------------------------------------------------------------
+EVALUATION (held-out)
+--------------------------------------------------------------------------------
+NAS in-domain held-out (11 wells, ~553 objects):
+  IoU 0.5 : precision 0.843  recall 0.886  F1 0.864  AP 0.802
+  IoU 0.6 : precision 0.819  recall 0.861  F1 0.839
+  IoU 0.7 : precision 0.768  recall 0.808  F1 0.788
+  IoU 0.8 : precision 0.659  recall 0.692  F1 0.675
+  IoU 0.9 : precision 0.311  recall 0.327  F1 0.319
+
+New-condition held-out (40 wells):
+  F1@0.5 0.885 | mAP(0.5-0.9) 0.830 | Aggregated Jaccard 0.920 | Dice 0.963
+
+Control (base cpsam_v2, no fine-tuning) on the same NAS held-out:
+  F1@0.5 0.713 | mAP 0.322 | Aggregated Jaccard 0.426
+  => fine-tuning raises F1 by +0.15 and boundary accuracy (AJI) by ~2x.
+
+Round 1 vs Round 2 (same code, same held-outs):
+  NAS held-out (in-domain) : F1 0.867 -> 0.864  (tie within noise on 11 wells)
+  New-condition held-out   : F1 0.863 -> 0.885 (+0.022); AJI 0.784 -> 0.920
+
+--------------------------------------------------------------------------------
+INTENDED USE / LIMITATIONS
+--------------------------------------------------------------------------------
+- Built to COUNT and SIZE Toxoplasma vacuoles per field/condition in fixed
+  fluorescence assays.
+- Strong on parasite-rich images; host-cell-marker images (tubulin/cellmask)
+  are the weakest domain and are the focus of ongoing curation (round 3).
+- NAS held-out is only 11 wells; the new-condition held-out shares plates/markers
+  with its training wells (within-condition generalisation, not a fresh lab).
+- Not validated on live/brightfield imaging or on non-Toxoplasma parasites.
+
+--------------------------------------------------------------------------------
+FILES IN THIS FOLDER
+--------------------------------------------------------------------------------
+  cpsam_v2_toxo_r2            the model weights (load this)
+  round2.log                 round-2 training composition + baselines
+  round2_heldout_metrics.csv per-IoU metrics on the NAS held-out
+  round2_vs_round1.csv       round-1 vs round-2 head-to-head
+  train_report.json          round-1 config + metrics + the 11 held-out names
+  vanilla_vs_finetuned.json  base cpsam_v2 vs fine-tuned control
+
+LICENSE : (choose before upload — CC-BY-4.0 suggested for weights + data-derived)
+CITATION : Olafsson E., Carruthers Lab — Toxoplasma PV segmentation (CellposeSAM), 2026.