<p>Deep learning-based computed tomography (CT) image prediction models often suffer from poor local feature perception, severe class imbalance, and a scarcity of high-quality labeled data. Existing fine-tuning methods are sample-inefficient and prone to overfitting. To address these issues, this study proposes a Qwen3-VL vision-language model based on adaptive semi-supervised learning. A novel data augmentation strategy for minority classes is introduced, utilizing diffusion models to learn the distribution of minority-class data and generate new samples. Perceptual similarity and generation quality are evaluated to ensure visual fidelity, while two-level feature selection is performed based on local density to increase the number of high-quality samples in the feature space. A fine-tuning method based on log-sum regularization is proposed, which decomposes pre-trained weights into direction and magnitude components, guiding the magnitude along specific directions to enhance the perception of fine-grained features. The multi-view prediction problem is transformed into a mass function using evidence theory, quantifying uncertainty and generating high-confidence pseudo-labels to improve prediction accuracy and generalization performance. Experiments are conducted on 1 980 CT samples of interstitial lung disease related to rheumatoid arthritis from Guang’anmen Hospital, China Academy of Chinese Medical Sciences, using three-fold cross-validation. The proposed model achieves accuracy, recall, and F1-score of 89.85%, 89.92%, and 91.40%, respectively. Compared to state-of-the-art models, it demonstrates higher prediction accuracy and generalization performance.</p>