source_code/gliomasam3_moe/visualizations/README.md

Visualization Suite (no code changes)

This folder adds standalone visualization scripts. It does not modify any existing training/inference code.

Quick start

1. Prepare prediction folders for each method (ours + baselines).
   Supported formats (per-case):

   • *_label.nii.gz (label map, values {0,1,2,4})
   • *_regions_prob.nii.gz (probabilities for WT/TC/ET)
   • *_regions_bin.nii.gz (binary WT/TC/ET)
   • {case_id}.nii.gz (SegMamba 3-channel)

2. Copy and edit the config:

   cp visualizations/vis_config_example.yaml visualizations/vis_config.yaml
   

3. Run (full suite):

   python visualizations/vis_suite.py --config visualizations/vis_config.yaml --checkpoint /path/to/ckpt.pt --run all
   

4. Run a subset (comma-separated):

   python visualizations/vis_suite.py --config visualizations/vis_config.yaml --checkpoint /path/to/ckpt.pt --run qualitative,et_absent,boundary
   

Notes

• Model-dependent visualizations (et_absent, moe, dual_domain, ampmix) require --checkpoint.
• Aux outputs (pi_ET, MoE routing, spectral outputs) are cached to predictions.aux_dir when provided.
• Output figures are saved to visualization.output_dir (default: visualizations/outputs).
• If matplotlib is missing, install it in your environment.