This study employs Grad-CAM to enhance the interpretability of Deep Learning models for tumor segmentation on the Breast Ultrasound and LiTS17 datasets. We evaluated three segmentation models: U-Net, MultiResUNet, and DCUNet, using metrics such as accuracy, precision, recall, Intersection over Union (IoU), Dice Coefficient, and various loss functions. The results indicate that MultiResUNet consistently outperforms U-Net and DCUNet across both datasets. Grad-CAM heatmaps revealed that MultiResUNet exhibits a high degree of focus on tumor regions, leading to superior segmentation accuracy. In contrast, DCUNet and U-Net showed moderate to low focus and targeting accuracy. By providing visual explanations of model decisions, Grad-CAM enhances the transparency and trustworthiness of segmentation models, thereby increasing trust among medical experts in the decision-making process.