Within medical image segmentation, traditional deep convolutional neural networks often struggle to make accurate inferences due to domain shifts between source and target data. This issue is particularly pronounced in clinical settings, where the specialized and confidential nature of medical data leads to a scarcity of annotated information. Although numerous solutions exist, they are frequently constrained in clinical environments by limitations in data collection and computational complexity. To address the issue of domain shifts in scenarios where medical data is scarce, we propose a novel single-source domain generalization algorithm—combining Sharpness-Aware Minimization (SAM) with the Focal and Tversky loss strategy (SAM-FT)—aimed at enhancing the model’s inference capabilities across different domains. SAM reduces the risk of overfitting across diverse domains by optimizing the sharpness of the loss function during training. This method enhances the model’s ability to recognize rare lesions, thereby stabilizing its performance in unencountered domains. Additionally, by adjusting the weighting of positive and negative samples, Focal loss enables the model to prioritize difficult-to-segment lesion areas, ensuring enhanced attention to regions that are more susceptible to segmentation errors or recognition challenges. As a variant of Focal loss, the Tversky loss further optimizes the model’s capability to recognize lesions of varying sizes, with particularly significant improvements in the segmentation of small lesion areas. Extensive experimental validations on the Fundus and Prostate datasets against state-of-the-art models have confirmed the efficacy of our strategy in enhancing the accuracy of medical image segmentation and the generalization capabilities of the model.