The Protocol Genome: A Self-Supervised Learning Framework from DICOM Headers

In this paper, we propose the Protocol Genome, a self-supervised learning framework from DICOM headers, achieving AUROC 0.901 (vs 0.847 baseline) and ECE 0.036 (vs 0.058) on fully held-out external validation. Our method demonstrates significant improved calibration and robustness across multiple different modalities (CT, MRI, CXR) and vendors. Clinical imaging flows through PACS and DICOM, whose protocols (scanner make/model; sequence; reconstruction kernel; kVp; TR/TE; slice thickness) dictate contrast, noise, and artifact profiles. These protocol choices give rise to hidden confounders that prevent cross-site generalization of image-only neural networks and challenge multi-center deployment. We present The Protocol Genome, a self-supervised learning (SSL) framework where structured DICOM headers are treated as a genomic code, and protocol-aware yet clinically robust image representations are learned. The Protocol Genome extracts tokenized embeddings for de-identified DICOM header fields and matches these with image-related features through: (1) protocol–image contrastive learning, (2) masked protocol prediction, and (3) protocol–protocol translation across series. We experiment with 1.26M studies (7 health systems, 31 scanners from 3 vendors; CT, MR, CR/DR modalities) and evaluate across three downstream tasks: (A) chest CT triage for acute PE, (B) brain MRI low-grade vs. high-grade glioma classification, and (C) chest radiograph cardiomegaly detection. Compared to strong SSL baselines (SimCLR, MAE) and ImageNet transfer, Protocol Genome pretraining increases external-site AUROC by +0.046 (95% CI: +0.031–+0.060) for PE, +0.058 (+0.036–+0.079) for glioma, and +0.041 (+0.028–+0.054) for cardiomegaly; calibration (ECE) improves by 25–37%. Further DeLong tests support significance (all p<0.001). Ablations indicate gains remain with 10–20% labeled data. Clinically, the method is applicable to reducing false positives at protocol borders and can be integrated into a PACS (DICOM C-FIND/C-MOVE, DICOMweb QIDO/WADO). We release a model card and deployment recommendations, with de-identification and bias auditing steps.