Halder Arindam, Dalal Anogh, Gharami Sanghita, Wozniak Marcin, Ijaz Muhammad Fazal, Singh Pawan Kumar
Department of Information Technology, Jadavpur University, Jadavpur University Salt Lake Campus, Plot No. 8, Salt Lake Bypass, LB Block, Sector III, Salt Lake City, Kolkata, 700106, West Bengal, India.
Faculty of Applied Mathematics, Silesian University of Technology, Kaszubska 23, Gliwice, 44100, Poland.
Sci Rep. 2025 Feb 20;15(1):6268. doi: 10.1038/s41598-025-90423-3.
Skin cancer is widespread and can be potentially fatal. According to the World Health Organisation (WHO), it has been identified as a leading cause of mortality. It is essential to detect skin cancer early so that effective treatment can be provided at an initial stage. In this study, the widely-used HAM10000 dataset, containing high-resolution images of various skin lesions, is employed to train and evaluate. Our methodology for the HAM10000 dataset involves balancing the imbalanced dataset by augmenting images followed by splitting the dataset into train, test and validation set, preprocessing the images, training the individual models Xception, InceptionResNetV2 and MobileNetV2, and then combining their outputs using fuzzy logic to generate a final prediction. We examined the performance of the ensemble using standard metrics like classification accuracy, confusion matrix, etc. and achieved an impressive accuracy of 95.14% and the result demonstrates the effectiveness of our approach in accurately identifying skin cancer lesions. To further assess the efficiency of the model, additional tests have been performed on the DermaMNIST dataset from the MedMNISTv2 collection. The model performs well on the dataset and transcends the benchmark accuracy of 76.8%, achieving 78.25%. Thus the model is efficient for skin cancer classification, showcasing its potential for clinical applications.
皮肤癌很常见，且可能致命。根据世界卫生组织（WHO）的数据，它已被确认为主要死因之一。尽早发现皮肤癌至关重要，以便在初期就能提供有效的治疗。在本研究中，使用了广泛应用的HAM10000数据集，该数据集包含各种皮肤病变的高分辨率图像，用于训练和评估。我们针对HAM10000数据集的方法包括通过图像增强来平衡不平衡数据集，然后将数据集拆分为训练集、测试集和验证集，对图像进行预处理，训练单个模型Xception、InceptionResNetV2和MobileNetV2，然后使用模糊逻辑组合它们的输出以生成最终预测。我们使用分类准确率、混淆矩阵等标准指标检查了集成模型的性能，达到了令人印象深刻的95.14%的准确率，结果证明了我们的方法在准确识别皮肤癌病变方面的有效性。为了进一步评估模型的效率，我们对MedMNISTv2数据集中的DermaMNIST数据集进行了额外测试。该模型在该数据集上表现良好，超越了76.8%的基准准确率，达到了78.25%。因此，该模型在皮肤癌分类方面效率很高，展现了其在临床应用中的潜力。
