<![CDATA[The limited availability of labelled medical data is a real obstacle in building a supervised learning-based Radiology analysis system. This study offers an alternative approach, unlabeled multimodal clustering, that simultaneously combines clinical information sources: chest X-ray images and physician clinical reports. Visual features are extracted from 1,000 images using a pretrained ResNet152, and text features are constructed using a Dual TF-IDF Vectorizer that processed the findings and conclusions columns separately with explicit weightings (0.4 and 0.6), then enriched with 10 related topics from the Latent Dirichlet Allocation (LDA) model. These two modalities are combined through a weighted late fusion strategy after being normalized by L2, with a visual weight of 0.5 and a text weight of 0.5, respectively. Before clustering, dimensionality reduction is performed using UMAP with 50 components and a cosine metric. Determining the optimal K value using the Elbow and Silhouette Score methods in the range of K=2 to K=15 showed K=8 as the best choice with a Silhouette Score value of 0.7524. The clustering results reveal eight clinical groups reflecting diverse diagnostic patterns, ranging from normal to abnormal findings, such as cardiomegaly and pneumonia, that require clinical attention. This approach has the potential to serve as the basis for a medical data exploration system that relies on no label annotations.]]>
Copyrights © 2026
