Deep-learning background field removal (BFRnet): a 3D dual-frequency octave-convolution U-net trained to predict the background field of the brain — including brains with significant pathological susceptibility sources (haemorrhage, calcification). Consumes the total field (ppm) and predicts the background field; the local tissue field is total − background, masked. The authors' trained MATLAB network was exported to ONNX and is run here with ONNX Runtime — no MATLAB Runtime — reproducing the MATLAB output to ~1e-8 at a fraction of the memory and image size. QSM-CI reconstruction method bfrnet. Browse and run it at https://qsmxt.github.io/QSM-CI/submission.html?method=bfrnet.