Classification models with human-interpretable explanations improve clinicians’ trust and diagnostic usability. A promising approach is to integrate pathology-related visual attributes used by radiologists, aligning AI decisions with clinical reasoning. Radiologists use attributes such as shape and texture as established diagnostic criteria, and mirroring these in AI decision-making enhances transparency and enables explicit validation of model outputs. However, progress is limited by the scarcity of large medical image datasets annotated with such attributes. To address this, we propose generating attribute-annotated images using an attribute-conditional diffusion model [1]. Applied to the LIDC-IDRI dataset, which contains lung nodule images with malignancy and attribute labels the model produces realistic, high-quality images, validated through a radiologist user study. Nevertheless, the data-demanding nature of diffusion models leads to substantial performance degradation when only limited annotated data are available, especially with 20 images. We therefore introduce a semi-conditional training strategy that also leverages unlabeled images by dynamically enabling conditioning. This reduces annotation effort while maintaining reasonable image quality. The second step of this work is to incorporate the generated images into the training of explainable models, such as HierViT [2]. This improves attribute prediction accuracy by 13.4% and malignancy prediction accuracy by 1.8% compared with using only the small annotated dataset. These results demonstrate that domain-driven synthetic augmentation can meaningfully reduce annotation burden and make explainable medical image classifiers more applicable.