Lt Vit Cxr Eval

lt-vit-cxr-eval

LT-ViT: A Vision Transformer for multi-label Chest X-ray classification — Marikkar et al. (2023) (arXiv:2311.07263, 2023)

What this evaluates

Evaluates a vision transformer's ability to perform multi-label classification on chest X-ray images. It probes the model's capacity to detect multiple pathologies simultaneously and model inter-label dependencies using learnable label tokens.

Datasets

• NIH-CXR14 — total ?; splits: test (-1)
• CheXpert-5 — total ?; splits: test (-1)
• CheXpert-13 — total ?; splits: test (-1)

Metrics

• AUC (%) (primary) — range: percent
  • Area Under the Receiver Operating Characteristic Curve computed for multi-label classification. Typically calculated per pathology and averaged (macro-average) across all labels.

Input / output format

Input: Chest X-ray radiograph images resized to 224×224 pixels.

Output: Multi-label probability scores or binary predictions for each of the 5 or 13 specified pathologies.

Scoring recipe

def compute_auc(predictions, gold):
    # predictions: (N, C) float array of probabilities
    # gold: (N, C) binary array
    aucs = []
    for c in range(predictions.shape[1]):
        auc = roc_auc_score(gold[:, c], predictions[:, c])
        aucs.append(auc)
    return np.mean(aucs) * 100

Common pitfalls

• AUC is reported as a percentage in tables, but the underlying metric ranges from 0 to 1.
• Multi-label AUC averaging strategy (macro vs. micro) is not explicitly specified in the text.
• Image resolution is fixed at 224×224 for fair comparison; deviating from this changes performance significantly.

Evidence (verbatim from paper)

Table 1 presents the comparison between the proposed method and previously established multi-label classification methods using transformers.

Table 1: Comparison of LT-ViT vs. existing benchmarks pre-trained on NIH-CXR14.

Citation

@misc{marikkar2023ltvit,
  title={LT-ViT: A Vision Transformer for multi-label Chest X-ray classification},
  author={Marikkar et al. (2023)},
  year={2023},
  note={arXiv:2311.07263}
}

• arXiv: 2311.07263