reversely/tcga-ucec-embedding

TCGA-UCEC endometrial pathology: page images + ColPali embeddings

Read-only serving assets for an endometrial pathology report review tool, derived from the public TCGA / GDC endometrial (UCEC) corpus. All source data is public and de-identified; embeddings come from the open-weight ColPali model.

Contents

• <patient_filename>/page_NNN.png -- rendered report page images at 200 DPI (545 cases, 1921 pages), one folder per case.
• <patient_filename>.npz -- ColPali page embeddings for that case: one float16 array per page, keyed page_000, page_001, ...
• query_embeddings.npz -- ColPali embeddings for the 9 CAP/ICCR field queries (histologic_type, histologic_grade, myometrial_invasion_percent, cervical_stromal_invasion, lymphovascular_space_invasion, margins, regional_lymph_node_status, pathologic_stage, molecular_classification).
• ocr/TCGA_Reports.csv -- OCR text for the reports.

Page and query vectors are scored with late-interaction MaxSim (ColBERT-style) on CPU, no GPU needed at serve time.

Provenance

• Page images: GDC pathology-report PDFs (files-API lookup by case barcode + data_type="Pathology Report"), rendered to PNG at 200 DPI (PyMuPDF).
• OCR text: tatonetti-lab/tcga-path-reports (Kefeli et al. 2024, Patterns).
• Embeddings: vidore/colpali-v1.3, multi-vector, bf16 inference, stored float16.