FERDINAND: An Open Framework for Download and AI-ready Preprocessing of IMPC 2D Radiographs

Abstract

The International Mouse Phenotyping Consortium (IMPC) maintains a large archive of whole-body dorsoventral radiographs. However, heterogeneity in file formats, metadata completeness, orientation, and image quality limits its reuse for large-scale quantitative analysis. We present FERDINAND (Framework for Extracting and Restructuring Distributed IMPC radiographs into a Normalised, AI-ready Navigable Dataset), an open framework for programmatic access, harmonisation, quality control, and curation of IMPC radiographs. FERDINAND combines metadata retrieval and local radiograph archiving with centre-aware preprocessing, automated exposure screening, orientation standardisation, and clustering-assisted visual curation. Applied to IMPC Data Release 24, FERDINAND organises 104,819 radiographs acquired between February 2012 and June 2025 across ten phenotyping centres and 6,306 single-gene knockout mouse lines, with semi-automated curation recorded for every image. As an independent validation of the harmonised output, a DeepLabCut-based bone-morphometry pipeline showed that FERDINAND-processed radiographs improve data efficiency for landmark detection: at reduced training-set sizes, they achieve better performance than original radiographs and require fewer manual annotations to reach a given level of quality. FERDINAND provides a reproducible foundation for artificial intelligence (AI)-ready analysis and scalable quantitative phenotyping of heterogeneous IMPC and comparable radiograph resources. The complete FERDINAND source code is openly available on GitHub.