Background/Objectives: Melanoma, an aggressive form of skin cancer, accounts for a significant proportion of skin-cancer-related deaths worldwide. Early and accurate differentiation between melanoma and benign melanocytic nevi is critical for improving survival rates but remains challenging because of diagnostic variability. Convolutional neural networks (CNNs) have shown promise in automating melanoma detection with accuracy comparable to expert dermatologists. This study evaluates and compares the performance of four CNN architectures—DenseNet121, ResNet50V2, NASNetMobile, and MobileNetV2—for the binary classification of dermoscopic images.Methods: A dataset of 8825 dermoscopic images from DermNet was standardized and divided into training (80%), validation (10%), and testing (10%) subsets. Image augmentation techniques were applied to enhance model generalizability. The CNN architectures were pre-trained on ImageNet and customized for binary classification. Models were trained using the Adam optimizer and evaluated based on accuracy, area under the receiver operating characteristic curve (AUC-ROC), inference time, and model size. The statistical significance of the differences was assessed using McNemar’s test.Results: DenseNet121 achieved the highest accuracy (92.30%) and an AUC of 0.951, while ResNet50V2 recorded the highest AUC (0.957). MobileNetV2 combined efficiency with competitive performance, achieving a 92.19% accuracy, the smallest model size (9.89 MB), and the fastest inference time (23.46 ms). NASNetMobile, despite its compact size, had a slower inference time (108.67 ms), and slightly lower accuracy (90.94%). Performance differences among the models were statistically significant (p< 0.0001).Conclusions: DenseNet121 demonstrated a superior diagnostic performance, while MobileNetV2 provided the most efficient solution for deployment in resource-constrained settings. The CNNs show substantial potential for improving melanoma detection in clinical and mobile applications.
**背景/目的：** 黑色素瘤是一种侵袭性皮肤癌，在全球皮肤癌相关死亡中占很大比例。早期准确区分黑色素瘤与良性黑素细胞痣对于提高生存率至关重要，但由于诊断存在差异性，这仍然具有挑战性。卷积神经网络在自动化黑色素瘤检测方面展现出潜力，其准确度可与皮肤科专家相媲美。本研究评估并比较了四种CNN架构——DenseNet121、ResNet50V2、NASNetMobile和MobileNetV2——在皮肤镜图像二分类任务中的性能。 **方法：** 研究使用来自DermNet的8825张皮肤镜图像数据集，将其标准化并划分为训练集（80%）、验证集（10%）和测试集（10%）。应用图像增强技术以提高模型的泛化能力。所有CNN架构均在ImageNet上进行预训练，并针对二分类任务进行定制化调整。模型使用Adam优化器进行训练，并基于准确率、受试者工作特征曲线下面积、推理时间和模型大小进行评估。使用McNemar检验评估模型间性能差异的统计学显著性。 **结果：** DenseNet121取得了最高的准确率（92.30%）和AUC值（0.951），而ResNet50V2则记录了最高的AUC值（0.957）。MobileNetV2在效率与性能之间取得了良好平衡，实现了92.19%的准确率、最小的模型大小（9.89 MB）以及最快的推理时间（23.46毫秒）。NASNetMobile尽管模型尺寸紧凑，但推理时间较慢（108.67毫秒），准确率也略低（90.94%）。模型间的性能差异具有统计学显著性（p < 0.0001）。 **结论：** DenseNet121展现出卓越的诊断性能，而MobileNetV2则为资源受限环境下的部署提供了最高效的解决方案。这些卷积神经网络在改善临床和移动应用中的黑色素瘤检测方面显示出巨大潜力。
肿瘤(癌症)患者之家