Adaptive clustering for medical image analysis using the improved separation index.

Area of Science:

• Biomedical data analysis
• Machine learning
• Medical imaging and signal processing

Background:

• Clustering high-dimensional biomedical data without knowing the number of clusters is a significant challenge.
• Existing methods often require parameter tuning or prior knowledge, limiting their applicability.

Purpose of the Study:

• To introduce SONSC (Separation-Optimized Number of Smart Clusters), an adaptive and interpretable clustering framework.
• To automatically determine the optimal number of clusters using a novel internal validity metric, the Improved Separation Index (ISI).

Main Methods:

• SONSC iteratively maximizes the ISI, which assesses intra-cluster compactness and inter-cluster separability.
• The framework operates without supervision or parameter tuning, offering an automated approach.
• Evaluated on benchmark datasets (MNIST, CIFAR-10) and clinical data (chest X-ray, ECG, RNA-seq).

Main Results:

• SONSC consistently outperformed K-Means, DBSCAN, and spectral clustering on ISI, Silhouette score, and normalized mutual information (NMI).
• Identified clinically relevant structures in real-world biomedical data, aligning with expert-labeled categories.
• Demonstrated superior performance in both numerical metrics and clinical interpretability.

Conclusions:

• SONSC offers a robust, scalable, and trustworthy solution for unsupervised biomedical data analysis.
• The framework's ability to unify algorithmic performance with medical interpretability supports its use in diagnostic and patient stratification pipelines.
• Provides a significant advancement in analyzing complex, high-dimensional biomedical datasets.