DEEP LEARNING-ENABLED EARLY DIAGNOSIS OF SKIN CANCER

Abstract

Using a publicly accessible Kaggle dataset taken from the ISIC Archive, we assessed three high-performance convolutional neural network architectures: ResNet-50, EfficientNet-B0, and Inception-V3. A standardised workflow including data normalisation, training-validation splitting, convolutional identification of features, and optimisation using gradient descent and cross-entropy loss was used to train these models using preprocessed dermoscopic pictures. Comparing these models' performances and evaluating their efficacy in classifying skin lesions—specifically, in differentiating amongst benign and malignant cases—was the primary objective of the study. According to the trial findings, Inception-V3 outperformed EfficientNet-B0 and ResNet-50, which had respective classification accuracy of 95.1% and 93.5%, with the highest classification accuracy of 98.6%. Along with its high accuracy, Inception-V3 also performed better in every other statistical indicator, such as F1-score (98.4%), recall (98.5%), and precision (98.2%). An important addition of this study is the 9% performance boost in classification accuracy compared to the baseline models. EfficientNet-B0 achieved 94.2% recall and a 95.8% accuracy rate, demonstrating a solid balance between efficiency and performance. With an F1-score of 90.8%, ResNet-50 generated competitive performance while trailing somewhat. Confusion matrix research demonstrates Inception-V3's potential for clinical implementation by confirming its excellent accuracy and low rate of incorrectly classifying malignant patients. All things considered, this work shows that deep learning models—in particular, Inception-V3—are quite successful in automatically detecting skin cancer.