New study benchmarks 5 clustering algorithms for medical image compression
Agglomerative clustering beats K-means on MRI and ultrasound by preserving fine diagnostic details.
A new arXiv paper (2607.09821) by Hamlomo and Atemkeng benchmarks clustering algorithms for medical image analysis, aiming to improve image compression without losing diagnostic details. The study compares k-means, mini-batch k-means, agglomerative hierarchical clustering, BIRCH, and bisecting k-means on MRI, ultrasound, and chest X-ray datasets. Using random search for hyperparameter optimisation, they assess cluster quality via Silhouette score, Davies-Bouldin index, and Calinski-Harabasz index.
Results show that while standard k-means and bisecting k-means achieve strong cluster cohesion, they often create few clusters with high intra-cluster variability, reducing effectiveness for adaptive compression. Agglomerative hierarchical clustering yielded the best intra-cluster homogeneity for MRI and ultrasound, making it ideal for preserving fine diagnostic features. For chest X-rays, mini-batch k-means delivered the best balance between clustering quality and compactness. BIRCH consistently lagged behind, raising questions about its suitability for medical imaging tasks.
- Agglomerative clustering outperformed all methods on MRI and ultrasound for preserving diagnostic details via intra-cluster homogeneity.
- Mini-batch k-means achieved the best balance of clustering quality and compactness specifically for chest X-ray images.
- BIRCH consistently underperformed across all three medical imaging modalities (MRI, ultrasound, chest X-ray).
Why It Matters
Better clustering algorithms can shrink medical image storage and transmission costs by 30–50% while keeping diagnostic accuracy intact.