Authors
Department of Medical Instruments Engineering Techniques, Bilad Alrafidain University College, 32001, Diyala, Iraq
[email protected]
Abstract
Unsupervised image segmentation remains one of the most persistent challenges in computer vision, particularly in fields lacking annotated data such as medical diagnostics and environmental monitoring. This paper introduces a novel segmentation model built on a modified U-Net backbone enhanced with self-supervised deep learning, dual spatial alignment (local and global), and explainability mechanisms including Grad-CAM and SLIC superpixels. The proposed framework was evaluated across three benchmark datasets from diverse domains: HyperKvasir (gastrointestinal endoscopy), PASCAL VOC 2012 (natural scenes), and ISIC 2018 (skin lesion images).Experimental results demonstrated robust segmentation outcomes, achieving DSC = 0.716 and Recall = 0.783, which outperform traditional unsupervised baselines. These findings were further validated through comparison with five recent methods, showing superior generalization and transparency. Additionally, the framework was successfully deployed in a practical application involving drought monitoring in Kirkuk, Iraq, by leveraging satellite imagery and unsupervised segmentation to support early warning systems. Overall, the results highlight the flexibility, interpretability, and domain adaptability of the proposed model, making it a promising tool for critical tasks in both medical and environmental domains.