Validation: tutorial 1 end to end (Python port vs MATLAB)

Run on 2026-10-30 with upstream Monalisa 0cb5cf9, MATLAB R2024b and the tutorial 1 data (3D radial phyllotaxis brain scan, 42 channels, Siemens raw data).

• MATLAB: tools/matlab_capture/scenarios/scenario_tutorial_recon.m

• Python: examples/tutorial_1/tutorial_1.py

• Comparison: tests/e2e/test_tutorial_1.py (monalisa.testing.metrics)

Both pipelines run coil sensitivity estimation (automatic mask, 5 refinement iterations), binning (all lines and 5 s sequential windows), mitosius preparation (normalization with the same fixed ROI in place of the interactive roipoly), and gridded reconstructions with mathilda (bmMathilda).

Output	relative L2	NMSE	SSIM (magnitude)
Binning masks (all lines, sequential 5 s)	identical
Coil sensitivity maps C (48³ × 42)	9.7e-05	9.3e-09	0.999999996
Volume elements (3D Voronoi, all lines)	4.1e-14
Gridded image used for normalization (80³)	6.7e-05	4.6e-09	0.999999995
Normalization value	6.1e-05 (relative difference)
x0: gridded reconstruction, all lines (80³)	3.8e-05	1.5e-09	0.999999998
x1: sequential bin 1 (80³)	4.1e-05	1.7e-09	0.999999998
x1: sequential bin 2 (80³)	4.1e-05	1.7e-09	0.999999998
x1: sequential bin 3 (80³)	4.1e-05	1.7e-09	0.999999998
x1: sequential bin 4 (80³)	4.1e-05	1.7e-09	0.999999998
x1: sequential bin 5 (80³)	4.1e-05	1.7e-09	0.999999998
x1: sequential bin 6 (80³)	4.1e-05	1.7e-09	0.999999998
x1: sequential bin 7 (80³)	4.1e-05	1.7e-09	0.999999998
x1: sequential bin 8 (80³)	4.1e-05	1.7e-09	0.999999998

All differences are at the level of single-precision rounding (the whole pipeline computes in single precision, like MATLAB). For reference, the previous port (2026 master project) reported relative L2 = 4.3 and SSIM = 0.18 on x0 and 0.14 / 0.93 on C.

Runtime on the development workstation (36 cores, 125 GB): Python 8.5 min, peak memory 35 GB; MATLAB 11.7 min (including the recording overhead of the capture harness).

↑