This study explores the efficacy of the Xception architecture for unsupervised image classification within a medical imaging context, specifically focusing on the challenge of classifying fundoscopy images from a Kaggle dataset into multiple stages of diabetic retinopathy. The results are analyzed using various metrics, including accuracy, sensitivity, and precision across different class combinations, revealing insights into the model’s ability to discern subtle differences between disease stages. The confusion matrices derived from these classifications provide a visual representation of the model’s performance, highlighting its strengths and limitations in different configurations. Notably, the binary classification combination (012-34) achieved the highest accuracy (89.74%) and sensitivity (88.54%), indicating a more effective model behavior in grouping similar disease stages.