Accurate segmentation of cardiac chambers from echocardiography is essential for assessing heart function. However, this task remains challenging due to blurred boundaries and the complex structure of the heart. While the Segment Anything Model (SAM) is promising, it requires manual prompts and performs poorly in capturing the heart’s complex structures and indistinct boundaries. To address these limitations, we propose MSPD-SAM, a prompt-free framework for fully automatic cardiac segmentation. Specifically, our method enhances SAM’s encoder with a multi-scale adapter to capture the complex anatomical structures of the heart. In addition, we propose a Parallel Mask Decoder to address the inherent trade-off in cardiac ultrasound segmentation between semantic coherence and boundary precision. We integrate these components into the MSPD-SAM framework and extensively evaluate our method on the public CAMUS and EchoNet-Dynamic datasets, achieving state-of-the-art performance and outperforming leading CNNs, Transformers, and other SAM-based adaptations.