You can deploy this model and start earning money today!

IshaChhikara/
cx-uni-abmil-21cohort

Name: cx-uni-abmil-21cohort
Availability: InStock

Views

No views yet

cx-uni-abmil-21cohort

license: mit library_name: pytorch tags:

medical-imaging
histopathology
whole-slide-image
chromosomal-instability
copy-number-signatures
attention-mil pipeline_tag: image-classification language:
en

A Gated Attention-MIL head that predicts the 17 Drews 2022 chromosomal-instability signatures (CX1–CX17) from whole-slide H&E pathology images, using frozen MahmoodLab/UNI tile embeddings.

This is the largest of the three cohort variants in the cx-histology series: 21 TCGA cohorts, fold-parallel trained across 5 GPUs. It supersedes cx-uni-abmil-15cohort and cx-uni-abmil-7cohort for the most cohort-diverse training distribution.

Headline result

Mean active ρ = +0.106 across 14 active signatures
0 H1 hits (ρ ≥ 0.30)
Top 5 fold-mean: CX1 +0.172, CX5 +0.166, CX12 +0.147, CX3 +0.130, CX10 +0.107
Trained on the 21-cohort superset (15-cohort cohorts + BRCA/LGG/GBM/LUAD/LUSC/HNSC)
Hyperparameters from the n=1,454 sweep winner: hidden=512, dropout=0.1, lr=1e-4

Architecture

WSI -> tile (224 px @ 0.5 mpp, Macenko-normalized)
    -> UNI (frozen) -> (N tiles, 1024)
    -> GatedABMIL (this model) -> (17,) CX activity prediction

Quick start

python

from huggingface_hub import snapshot_download
from modeling_cx_abmil import CXEnsemble
local = snapshot_download("IshaChhikara/cx-uni-abmil-21cohort")
model = CXEnsemble.from_pretrained(local, device="cpu")
cx = model.predict(my_uni_embeddings)  # (B, N, 1024) -> (B, 17)

External validation on CPTAC

Held-out test on 283 CPTAC-3 patients with paired UNI tile embeddings and recomputed CX activities (CINSignatureQuantification v1.2.0 on GDC ASCAT2 seg files, build hg38, DCIN=10) meets the pre-registered acceptance criterion |ρ_TCGA − ρ_CPTAC| ≤ 0.15 for all 14 active signatures:

Signature	ρ_TCGA	ρ_CPTAC	\|Δρ\|
CX6	+0.056	+0.058	0.002
CX3	+0.114	+0.102	0.012
CX9	+0.026	+0.010	0.017
CX8	+0.096	+0.068	0.027
CX2	+0.053	+0.013	0.040
CX12	+0.130	+0.062	0.069
CX4	+0.060	−0.014	0.074
CX1	+0.137	+0.056	0.081
CX11	+0.069	−0.019	0.088
CX17	+0.044	+0.133	0.090
CX16	+0.078	+0.176	0.098
CX13	+0.017	−0.097	0.114
CX10	+0.101	−0.014	0.114
CX5	+0.147	+0.021	0.125

CX16 and CX17 generalise better on CPTAC than on TCGA. The remaining twelve signatures retain the sign and rough magnitude of the TCGA estimates, with absolute ρ_CPTAC values smaller than ρ_TCGA in ten of fourteen cases.

The cohort-scaling story (read together with the other two repos)

Variant	n	Mean ρ	H1 hits
7-cohort	289	+0.219	3
15-cohort	1,454	+0.138	0
21-cohort	3,855	+0.106	0

Smaller cohort sets achieve larger pan-cohort ρ partly because cancer-type heterogeneity is doing the work. As cohorts are added, the within-biology signal becomes the dominant explanation. The 21-cohort number is the most honest estimate of the H&E→CIN biology link.

Files

modeling_cx_abmil.py — model + CXEnsemble 5-fold ensemble wrapper
inference_example.py
config.json
fold0.pt … fold4.pt
LICENSE (MIT)

Citation

In preparation. Code: github.com/ishaschhikara316/cx-histology.