Recent advances in medical imaging have intensified the demand for robust analysis methods to support diagnosis, treatment planning, and monitoring. However, current approaches still face significant challenges in two critical limitations: i) scarcity of annotated data, ii) poor model interpretability. To address these challenges, we propose MedCDA, a zero-shot segmentation-driven counterfactual data augmentation framework. Our approach introduces: (1) a boundary-aware gradient attention mechanism that sharpens focus on target boundaries via automatically simulated click points; (2) systematic counterfactual generation that removes lesions while preserving healthy semantics, enhancing diversity and reducing spurious correlations; and (3) multimodal large model integration for vision-language alignment, paired with a weighted loss fine-tuning strategy to improve classification robustness. Quantitative results demonstrate consistent improvements over state-of-the-art methods on three public benchmarks: dermoscopic image dataset HAM10000, breast ultrasound image dataset BUSI, and mammography dataset CBIS-DDSM. Ablation studies validate the effectiveness of each module. MedCDA establishes a novel “segmentation-augmentation-reasoning” paradigm, offering an extensible framework for other healthcare decision-making scenarios and providing new solutions for medical image analysis.