wenmin-wu/cv-per-modality-ensemble-averaging

Per-Modality Ensemble Averaging

Overview

Multi-modal medical imaging (MRI sequences, CT windows) captures complementary information. Rather than concatenating modalities into a single multi-channel input, train an independent model per modality and average their sigmoid/softmax outputs at inference. This avoids missing-modality issues and lets each model specialize, often outperforming multi-channel approaches when modality quality varies.

Quick Start

import numpy as np
from sklearn.metrics import roc_auc_score

def train_per_modality(train_df, val_df, modalities, train_fn, predict_fn):
    """Train one model per modality, return ensemble predictions."""
    models = {}
    for mod in modalities:
        models[mod] = train_fn(train_df, val_df, modality=mod)

    val_preds = np.zeros(len(val_df))
    for mod in modalities:
        val_preds += predict_fn(models[mod], val_df, modality=mod)
    val_preds /= len(modalities)

    auc = roc_auc_score(val_df['label'], val_preds)
    return models, auc

modalities = ['FLAIR', 'T1w', 'T1wCE', 'T2w']
models, auc = train_per_modality(df_train, df_val, modalities, train_fn, predict_fn)

Workflow

1. Identify available imaging modalities per patient
2. Train one model per modality using modality-specific data
3. At inference, run each model on its corresponding modality
4. Average predictions across all modalities
5. Handle missing modalities by averaging only available ones

Key Decisions

• Equal weighting: simple average works well; learned weights add complexity for marginal gain
• Missing modalities: average over available ones only, don't zero-fill
• vs multi-channel: ensemble is more robust when some modalities have quality issues
• Model architecture: same architecture per modality, but weights are independent

References

• Efficientnet3D with one MRI type