Initial upload via tools/push_to_hf.py (architecture: ilex.models.bme_x.BMEX)

config.json

@@ -21,7 +21,7 @@
   "equinox_version": "0.13.8",
   "ilex_version": "0.0.0.dev0",
   "image_classes": "Single-channel 3D brain MRI (T1- or T2-weighted, per the variant's modality), resampled to 0.8 mm isotropic resolution per the upstream preprocessing, patched into 32^3 or 64^3 windows for inference.",
-  "intended_use": "Inference-only enhancement of a single-channel 3D brain MRI volume. Inputs are expected to be conformed to 0.8 mm isotropic, brain- extracted, and intensity-normalised via histogram-matching against an age-specific template (the upstream's BME_X_enhanced.py pipeline performs all three steps), then tiled into 32^3 patches with sliding-window inference at 0.85 overlap. Each variant's weights are tuned for a specific age range + modality (T1 vs T2). Per Sun et al. NBME 2025, the Month24 variants are the de-facto adult-applicable models. The seg head's 4-channel pre-softmax relu output (BG / WM / GM / CSF; apply softmax for probabilities) is an auxiliary signal that drives the reconstruction head's tissue-aware enhancement; the recon head's 1-channel output is the enhanced volume (motion- corrected, denoised, super-resolved, harmonized in one forward). CRITICAL CAVEAT (v0) -- without the upstream's histogram-matching preprocessing, the recon head produces all-zero output for at least eight of the fifteen released variants, including both adult-applicable ones (month24-t1, month24-t2). The seg head's argmax also collapses to all-background on un-preprocessed inputs. This is a property of the upstream's trained weights, not a port bug -- PyTorch and JAX agree to ~5e-5 max abs in all measured cases. See DEVNOTES.md \"Recon head usability\" for the per-variant breakdown. Until the BMEX image-domain pipeline lands in ilex.models.bme_x.pipeline, treat this port as the seg head + a partial recon head; do not consume recon directly without running the upstream's BME_X.py preprocessing first.",
+  "intended_use": "Inference-only enhancement of a single-channel 3D brain MRI volume. Inputs are expected to be conformed to 0.8 mm isotropic, brain- extracted, and intensity-normalised via histogram-matching against an age-specific template (the upstream's BME_X_enhanced.py pipeline performs all three steps), then tiled into 32^3 patches with sliding-window inference at 0.85 overlap. Each variant's weights are tuned for a specific age range + modality (T1 vs T2). Per Sun et al. NBME 2025, the Month24 variants are the de-facto adult-applicable models. The seg head's 4-channel pre-softmax relu output (BG / WM / GM / CSF; apply softmax for probabilities) is an auxiliary signal that drives the reconstruction head's tissue-aware enhancement; the recon head's 1-channel output is the enhanced volume (motion- corrected, denoised, super-resolved, harmonized in one forward). IMPORTANT input contract -- BME-X expects input at the cohort- specific template's raw intensity scale (0-1000), NOT naively renormalised to [0, 1]. The v0.1 pipeline (ilex.models.bme_x.pipeline.enhance) handles this: it fetches the cohort template lazily from OSF via lytemaps.fetch_bmex, histogram- matches the input, tiles into 32^3 patches with 0.85 overlap, and aggregates the recon outputs. Consuming BMEX.__call__ directly on naively-normalised input produces all-zero recon output for about half the variants (including both adult-applicable ones) -- this is the input-contract mismatch, not a broken model (PyTorch and JAX agree to ~5e-5 in all measured cases). End-to-end inference on a real adult T1 patch (cvs_avg35 MNI, central 64^3) through pipeline.enhance with cohort='adult', suffix='T1w' produces recon range [0, 533.6] with 58.8% nonzero voxels and a diverse 4-class tissue segmentation; the pipeline is the validated usability surface.",
   "jax_version": "0.10.0",
   "network_data_format": {
     "inputs": {},