A Dataset-Driven AutoML Framework for Chest X-Ray Image Clustering
IEEE ACCESS, cilt.14, ss.96020-96044, 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 14
- Basım Tarihi: 2026
- Doi Numarası: 10.1109/access.2026.3705522
- Dergi Adı: IEEE ACCESS
- Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Compendex, INSPEC, Directory of Open Access Journals
- Sayfa Sayıları: ss.96020-96044
- Ondokuz Mayıs Üniversitesi Adresli: Evet
Özet
Automated machine learning (AutoML) has emerged as a powerful paradigm in supervised learning, where labeled data provide a clear basis for model selection and hyperparameter optimization. In contrast, its application to clustering tasks, particularly in medical imaging, remains relatively underexplored. This is largely due to the scarcity of reliable annotations and the difficulty of defining suitable optimization objectives in the absence of labels. In this work, we propose a dataset-driven AutoML framework for clustering chest X-ray images. The framework is designed to adapt to the intrinsic structure of the data without relying on manual labeling during model selection. To achieve this, we leverage self-supervised representations extracted from a Vision Transformer (ViT) pretrained using Self-Distillation with No Labels (DINO), which provide semantically meaningful embeddings that capture underlying visual patterns in the images. These representations are integrated into an AutoML pipeline formulated as a Combined Algorithm Selection and Hyperparameter optimization (CASH) problem. Importantly, the optimization process is guided exclusively by internal clustering validity indices, ensuring a fully unsupervised model selection procedure. External agreement measures are introduced only at the final evaluation stage, allowing for an interpretable assessment of how well the discovered clusters align with clinically meaningful categories. On the primary binary chest X-ray dataset (Normal vs. Pneumonia), the proposed framework automatically identifies an OPTICS-based clustering configuration, achieving a post-hoc classification-style test accuracy of 85.7%. The resulting clusters exhibit strong geometric structure, with a Silhouette score of 0.742, a Calinski-Harabasz index of 2529.14, and a Davies-Bouldin score of 0.408. In addition, the clustering demonstrates meaningful alignment with clinical labels, as reflected by a Rand Index of 0.754 and an Adjusted Rand Index of 0.508. Experiments on both binary and multi-class chest X-ray datasets further demonstrate that the proposed framework is capable of discovering geometrically coherent cluster structures that show meaningful, though moderate, alignment with reference clinical labels, even in the absence of labels during model selection. Overall, this study shows that combining self-supervised representation learning with internal-metric-driven AutoML provides a principled and data-driven approach for exploratory analysis in label-scarce medical imaging settings. The proposed framework reduces reliance on manual design choices while preserving interpretability, making it a promising strategy for exploratory analysis in annotation-limited medical imaging settings.