IU-Net-DDPC: A Novel Integrated U-Net and Dynamic Density Peaks Clustering Framework for Real-Time Medical Image Stream Analysis

The rapid expansion of Internet of Medical Things (IoMT) devices has transformed modern healthcare systems, generating continuous, unprecedented streams of unlabeled medical images in real time from multiple sources. Current medical image clustering methods suffer from important weaknesses in both computational efficiency and accuracy, and fail to adapt their decomposition parameters to the more challenging scenario of ever-changing feature distributions found in most clinical environments. Therefore, this paper proposes a novel framework termed “IU-NET-DDPC” that integrates a U-Net convolutional autoencoder for layered deep feature extraction and dynamic density peaks clustering (DDPC) to autonomously identify clusters without a predefined number of clusters. Instead of pairwise computation, it is efficient to compute for streaming batches using a KNN-based density estimation strategy. An adaptive split-and-merge approach, as well as delete procedures, in a dynamic cluster maintenance mechanism, are applied to changes caused by data distribution. Evaluation on a Cardiac Catheterization dataset at 200 training epochs and 0.2 dropout rate achieves 89.60% accuracy, NMI of 0.5182, ARI of 0.6268, Purity of 89.60%, F1-Score of 89.60%, and Jaccard index of 0.8115. IU-Net-DDPC outperformed various baseline methods across all metrics in comparative experiments. The practical applicability of the proposed framework confirms the suitability for real-time clinical IoMT deployment.