rphammonds/scalesurfer-stats

ScaleSurfer v567 (combined FreeSurfer 5/6/7) Stats Prediction Model

This repository contains a ScaleSurfer multi-head model for predicting FreeSurfer-style .stats targets from a T1w image and an aparc+aseg segmentation.

The v567 stats model uses the frozen combined ScaleSurfer v567 encoder and is trained on all available stats targets across FreeSurfer versions. The encoder itself starts from the original cross-entropy ScaleSurfer checkpoint and was continued with the FastSurferCNN v2.5.4-style soft-Dice plus weighted cross-entropy objective. Only the stats heads are initialized from the earlier all-data stats model; the legacy encoder weights are not loaded.

Files

• stats_model.safetensors: model weights and inference metadata in safetensors format.
• config.json: architecture and feature-schema metadata needed by ScaleSurferStatsPredictor.
• metadata.json: checksums and training/evaluation metadata.
• summary.csv, history.csv, target_metrics.csv: copied training diagnostics.

Test Summary

Group	Targets	Values	Normalized MAE	Median absolute percent error
aseg	119	39699	0.41874403593829795	6.72153377532959
global	64	16070	0.15233277114068536	1.0538854598999023
lh_aparc	306	106742	0.4084990393382036	7.2162065505981445
rh_aparc	306	106738	0.4122215262526054	7.313578844070435

Loading

from scalesurfer.stats import ScaleSurferStatsPredictor

predictor = ScaleSurferStatsPredictor.from_pretrained()
features = predictor.predict_subjects(subjects_dir, subjects, return_format="wide")

This model is intended for research workflows and is not a clinical diagnostic device.