Mvsl Biomedical Fewshot Eval

mvsl-biomedical-fewshot-eval

Multi-View Synergistic Learning with Vision-Language Adaption for Low-Resource Biomedical Image Classification — Xiaoliu Luo et al. (2026) (arXiv:2604.23977, 2026)

What this evaluates

Evaluates a vision-language model's few-shot classification capability on diverse biomedical images, testing cross-modal alignment, generalization to unseen disease categories, and robustness across multiple imaging modalities and anatomical regions.

Datasets

• CTKidney — total ?; splits: few-shot support/query (-1)
• DermaMNIST — total ?; splits: few-shot support/query (-1)
• Kvasir — total ?; splits: few-shot support/query (-1)
• RETINA — total ?; splits: few-shot support/query (-1)
• LC25000 — total ?; splits: few-shot support/query (-1)
• CHMNIST — total ?; splits: few-shot support/query (-1)
• BTMRI — total ?; splits: few-shot support/query (-1)
• OCTMNIST — total ?; splits: few-shot support/query (-1)
• BUSI — total ?; splits: few-shot support/query (-1)
• COVID-QU-Ex — total ?; splits: few-shot support/query (-1)
• KneeXray — total ?; splits: few-shot support/query (-1)

Metrics

• classification accuracy (%) (primary) — range: percent
  • Percentage of correctly predicted query images out of total query images in the few-shot support/query split.
• harmonic mean (HM) of base and novel accuracy — range: percent
  • HM = 2 * (base_acc * novel_acc) / (base_acc + novel_acc), where base_acc and novel_acc are classification accuracies on seen and unseen classes respectively.

Input / output format

Input: Biomedical images paired with disease category labels. For few-shot evaluation, models receive a support set of K labeled examples per class (K∈{1,2,4,8,16}) and 50 LLM-generated textual prompts initialized with 'a photo of a'.

Output: Predicted disease category label for each query image.

Scoring recipe

# Few-shot accuracy
support_set = sample_K_shots_per_class(K)
query_preds = model.predict(query_images, support_set)
correct = sum(1 for p, g in zip(query_preds, query_labels) if p == g)
accuracy = (correct / len(query_labels)) * 100

# Base-to-Novel HM
base_acc = (sum(1 for p, g in zip(base_preds, base_golds) if p == g) / len(base_golds)) * 100
novel_acc = (sum(1 for p, g in zip(novel_preds, novel_golds) if p == g) / len(novel_golds)) * 100
hm = 2 * (base_acc * novel_acc) / (base_acc + novel_acc)

Common pitfalls

• Results are averaged over three randomly sampled support sets per dataset, not a fixed train/val/test split.
• Harmonic mean (HM) can mask poor novel-class performance if base-class accuracy is high.
• Few-shot results vary significantly based on the specific K-shot support set sampling strategy.

Evidence (verbatim from paper)

Performance is quantified using the harmonic mean (HM) of accuracy across base and novel classes, measuring the model’s capability to recognize unseen disease categories while retaining knowledge of seen classes. Each result is reported as mean, calculated over three randomly sampled support sets per dataset.

Citation

@misc{luo2026mvsl,
  title={Multi-View Synergistic Learning with Vision-Language Adaption for Low-Resource Biomedical Image Classification},
  author={Xiaoliu Luo et al. (2026)},
  year={2026},
  note={arXiv:2604.23977}
}

• arXiv: 2604.23977