Radiological Metadata and Deep Image Feature Analysis for Breast Cancer Classification: An Explainable Multimodal Evaluation on CBIS-DDSM

Controlled multimodal ablation study

The study compares image only, clinical metadata only, and fusion models on the same CBIS DDSM data split. This setup measures the diagnostic value of each modality under equal conditions. The authors identify this type of controlled comparison as limited in earlier research.

Clinical metadata value

Radiologist defined features produce the highest performance. Mass shape, margin, assessment score, and subtlety reach a ROC AUC of 0.8737. This result exceeds both the image model and the fusion models. The finding shows clinical descriptors contribute more than deep image features in this dataset.

Feature dimension and fusion performance

The study evaluates image feature vectors with 64 and 1,280 dimensions. The larger feature vector reduces fusion performance because the model overfits the limited dataset. This result offers a practical lesson for multimodal medical AI projects with small datasets.

Grad CAM explainability

Grad CAM visualizations show the EfficientNet B0 model focuses on relevant lesion areas instead of background regions. These results support model interpretation and clinical confidence, even though the image only model achieves lower diagnostic performance.

Transparent negative finding

Simple feature concatenation fails to outperform the metadata only model. This negative result provides useful evidence for future research. Researchers should test attention based, gated, or other advanced fusion methods instead of assuming multimodal fusion will improve performance.