RAISE: Retrieval-Augmented Image Search with Saliency Explanations for Medical Imaging

A framework for content-based image retrieval with built-in saliency explanations, evaluated on three public medical imaging benchmarks. The system integrates a CLIP encoder, FAISS index, and a novel similarity-guided Grad-CAM module. It was authored by Din Bandhu and last updated on May 20, 2026.

Use Cases

• Benchmarking retrieval performance of medical CBIR systems based on the reported Recall@5 and mAP scores.
• Developing explainable AI tools for radiology based on the similarity-driven Grad-CAM heatmap generation method.
• Studying the scalability of approximate nearest-neighbor search in clinical settings based on the sub-millisecond query latency analysis.
• Conducting ablation studies on retrieval pipeline components based on the framework's modular design.

Strengths

• Achieved high retrieval performance on three benchmarks, with Recall@5 scores of 0.91, 0.88, and 0.94.
• Includes a novel explanation module generating saliency heatmaps with IoU scores ranging from 0.47 to 0.59 against ground-truth lesions.
• Demonstrated sub-millisecond query latency scalability on up to one million images.
• Framework components were systematically evaluated through an ablation study.

Limitations

• The dataset file size is only 14.3 KB, suggesting it likely contains code, model weights, or metadata rather than the full image data.
• Row count and specific column-level documentation are unknown, limiting suitability assessment for direct data analysis.
• The actual medical image data from the benchmarks (NIH ChestXray14, RSNA Pneumonia Detection, BraTS 2020) is not included in this package.

Provenance

Source

figshare

Collection Method

Likely contains the RAISE framework code and evaluation results, as described in the associated paper.

Freshness

Last updated 2026-05-20 06:55:59