Objective To investigate the effectiveness of a Mamba-enhanced convolutional and state-attention network (MECSA-Net) for classifying triple-negative breast cancer (TNBC) in ultrasound images. Methods A total of 1 059 breast ultrasound images were retrospectively collected, including 166 TNBC and 893 non-TNBC images. Data augmentation techniques were applied to mitigate class imbalance. A lightweight hybrid architecture MECSA-Net was proposed, featuring an extraction module named SAEffBlock (shuffle-aware efficient block), which integrated a state-space modeling branch (SSM-Branch) and a lightweight convolutional branch (EffConvBranch). Additionally, a multi-scale dilated fusion attention module (MDFA) was incorporated before the classifier to enhance the model’s ability to perceive multi-scale structures and model contextual information. Results MECSA-Net achieved an accuracy of 93.9%, a precision of 94.4%, an F1-score of 93.9%, and an area under curve of 0.976, outperforming mainstream models, including ResNet-18, ResNet-50, EfficientNet-B0, ViT-Base, and MedMamba-T. Confusion matrix analysis demonstrated high classification accuracy and low misclassification rates for both TNBC and non-TNBC samples. Ablation studies confirmed the crucial roles of both the EffConvBranch and MDFA modules in local texture representation and multi-scale structure discrimination, significantly enhancing classification performance and robustness. Conclusion MECSA-Net exhibits excellent accuracy and robustness in TNBC ultrasound image classification, indicating strong potential for clinical application and providing technical support for the intelligent preoperative diagnosis of TNBC.