Urundai Meeran Sabura Banu
Department of Electrical and Electronics Engineering, Saveetha Engineering College, Saveetha Nagar, Thandalam, Chennai: 602105, Tamilnadu, India.
Asian Pac J Cancer Prev. 2025 Aug 1;26(8):2919-2928. doi: 10.31557/APJCP.2025.26.8.2919.
The primary objective of this study is to enhance the accuracy and efficiency of binary skin lesion classification by optimizing the ResNet-50 convolutional neural network using the Whale Optimization Algorithm (WOA). This involves fine-tuning key hyperparameters such as learning rate, weights, and biases to improve predictive performance.
This study compares five CNN architectures: AlexNet, GoogleNet, VGG16, Resnet 50, and WOA-optimized Resnet 50. The dataset comprises 3,600 balanced images (224×244 resolution) of skin moles, evenly divided into 1,800 benign and 1,800 malignant cases. The models were trained on an open-access dermoscopic dataset to categorize skin lesions. WOA was applied to optimize Resnet 50's hyperparameters weight and bias learning rate. Model performance was analysed using accuracy, preci-sion, recall, F1 score, specificity, Matthews Correlation Coefficient (MCC), log loss, AUC-ROC, and infer-ence time. The confusion matrix was analyzed to assess misclassification rates. Result: The WOA-optimized Resnet 50 outperformed all other models, achieving 98.29% accuracy, higher than standard Resnet 50 (90.13%), GoogleNet (87.1%), AlexNet (86.53%), and VGG16 (81.18%). It also demonstrated superior recall (99.31%), specificity (97.07%), and an AUC-ROC of 99.84%, indicating excellent classification capability. The MCC score (0.9657) confirmed strong predictive reliability. Addi-tionally, the optimized model achieved the lowest log loss (0.0512), ensuring high confidence in predictions. With an inference time of 0.1488 seconds, it was significantly faster than standard Resnet 50 (1.029 seconds), making it computationally efficient. The confusion matrix confirmed its reliability, showing min-imal false positives (7) and false negatives (2).
WOA-optimized Resnet 50 significantly improves accuracy, recall, specificity, and computational efficiency for binary skin lesion classification. Compared to traditional deep learning models, it offers superior predictive performance while maintaining fast inference time. These findings suggest that WOA-enhanced deep learning can enhance dermatological diagnostics, aiding early detection and clinical decision-making. Future research may explore its application for multi-class skin lesion classification and real-time medical imaging systems.
本研究的主要目的是通过使用鲸鱼优化算法（WOA）优化ResNet-50卷积神经网络，提高二元皮肤病变分类的准确性和效率。这涉及微调关键超参数，如学习率、权重和偏差，以提高预测性能。
本研究比较了五种卷积神经网络架构：AlexNet、GoogleNet、VGG16、ResNet 50和WOA优化的ResNet 50。数据集包含3600张平衡的皮肤痣图像（分辨率为224×244），平均分为1800例良性和1800例恶性病例。这些模型在一个开放获取的皮肤镜数据集上进行训练，以对皮肤病变进行分类。应用WOA优化ResNet 50的超参数权重和偏差学习率。使用准确率、精确率、召回率、F1分数、特异性、马修斯相关系数（MCC）、对数损失、AUC-ROC和推理时间来分析模型性能。分析混淆矩阵以评估错误分类率。结果：WOA优化的ResNet 50优于所有其他模型，准确率达到98.29%，高于标准ResNet 50（90.13%）、GoogleNet（87.1%）、AlexNet（86.53%）和VGG16（81.18%）。它还表现出卓越的召回率（99.31%）、特异性（97.07%）和AUC-ROC为99.84%，表明具有出色的分类能力。MCC分数（0.9657）证实了强大的预测可靠性。此外，优化后的模型实现了最低的对数损失（0.0512），确保了对预测的高度信心。推理时间为0.1488秒，明显快于标准ResNet 50（1.029秒），使其计算效率高。混淆矩阵证实了其可靠性，显示出极少的假阳性（7例）和假阴性（2例）。
WOA优化的ResNet 50显著提高了二元皮肤病变分类的准确率、召回率、特异性和计算效率。与传统深度学习模型相比，它在保持快速推理时间的同时提供了卓越的预测性能。这些发现表明，WOA增强的深度学习可以改善皮肤病诊断，有助于早期检测和临床决策。未来的研究可以探索其在多类皮肤病变分类和实时医学成像系统中的应用。
