This paper investigates the optimization and performance evaluation of deep learning algorithms in medical image processing. Medical image processing is crucial for disease diagnosis and treatment planning, and deep learning has shown great potential in this field. However, challenges such as data scarcity, high labeling costs, and model interpretability remain. The study evaluates three deep learning models—U-Net, DeepLab, and DenseNet—on medical image classification tasks using ROC curves and AUC values. Results show that DenseNet outperforms the others in prediction accuracy and AUC value, while U-Net and DeepLab have slightly lower performance. The paper proposes an adaptive learning rate and momentum strategy to improve model training efficiency and performance. Experimental results on BRATS and LIDC-IDRI datasets demonstrate that DenseNet achieves the best performance, followed by U-Net, with DeepLab being less effective. The study highlights the importance of model optimization, data augmentation, and the need for further research into model structure improvements and multimodal data fusion. The findings suggest that DenseNet is suitable for high-precision segmentation, while U-Net is better for real-time applications. Future research should focus on enhancing model generalization and adaptability through diverse data sets and advanced techniques. The study provides valuable insights for improving medical image processing and diagnosis.This paper investigates the optimization and performance evaluation of deep learning algorithms in medical image processing. Medical image processing is crucial for disease diagnosis and treatment planning, and deep learning has shown great potential in this field. However, challenges such as data scarcity, high labeling costs, and model interpretability remain. The study evaluates three deep learning models—U-Net, DeepLab, and DenseNet—on medical image classification tasks using ROC curves and AUC values. Results show that DenseNet outperforms the others in prediction accuracy and AUC value, while U-Net and DeepLab have slightly lower performance. The paper proposes an adaptive learning rate and momentum strategy to improve model training efficiency and performance. Experimental results on BRATS and LIDC-IDRI datasets demonstrate that DenseNet achieves the best performance, followed by U-Net, with DeepLab being less effective. The study highlights the importance of model optimization, data augmentation, and the need for further research into model structure improvements and multimodal data fusion. The findings suggest that DenseNet is suitable for high-precision segmentation, while U-Net is better for real-time applications. Future research should focus on enhancing model generalization and adaptability through diverse data sets and advanced techniques. The study provides valuable insights for improving medical image processing and diagnosis.