Thyroid nodules, increasingly prevalent with age, require accurate assessment to identify malignancies. Ultrasound imaging, the primary diagnostic tool, has benefited from recent deep learning advancements. However, training models with imbalanced medical datasets poses challenges, as random data augmentation can introduce noise and distort critical features. This study employs weakly supervised data augmentation networks (WSDAN) and curriculum learning to improve thyroid nodule assessment using ultrasound images. WSDAN uses attention-guided data augmentation to enhance model generalization, while curriculum learning progressively exposes models to complex examples, improving their ability to differentiate benign from malignant nodules. Our approach, detailed through methodology and experimental results, shows that curriculum learning schematic II, where WSDAN first trains on easy images and progressively adds harder images, outperforms others. The results include precision (91%, 85%), recall (84%, 92%), and F1-score (87%, 88%) for benign and malignant classes, respectively, and an accuracy of 88%, demonstrating its efficacy in enhancing diagnostic accuracy and advancing automated thyroid nodule assessment.