PERBANDINGAN KINERJA DAN EFISIENSI ARSITEKTUR CNN VGG-16 DAN RESNET50 DALAM KANKER KULIT DATASET HAM10000

Abstrak

The primary focus of this research is to compare the capabilities and efficiency of two Convolutional Neural Network (CNN) architectures: ResNet50 and VGG16. This comparison is conducted for the task of classifying skin lesion images using the HAM10000 dataset. This dataset comprises dermatoscopic images representing various types of skin lesions, including melanoma, nevus, and other benign and malignant tumors. Both models were optimized using transfer learning with pretrained weights from ImageNet and trained with identical parameters. The research findings indicate that ResNet50 outperformed VGG16, achieving an accuracy of 80.96% on the testing data, whereas VGG16 only reached 74.85%. While ResNet50 demonstrated superior results in terms of validation and testing accuracy, as well as generalization capability, VGG16 performed better on majority classes. Both models encountered difficulties in recognizing minority classes, such as melanoma and dermatofibroma. This challenge is likely attributable to the imbalance in the number of samples across different classes. Although ResNet50 showed overall higher accuracy, this study also highlights the necessity for further approaches, such as data balancing and augmentation, to enhance performance on minority classes. This system holds significant potential to serve as a foundation for developing AI-based skin cancer detection applications, which could assist medical professionals in accelerating diagnoses and improving detection accuracy.

Identifier

• DOI: https://doi.org/10.23960/jitet.v14i2.9351
• OpenAlex: https://openalex.org/W7154022127