DEEP TRANSFER LEARNING FOR DERMOSCOPIC SKIN LESION CLASSIFICATION: BENCHMARKING XCEPTION, INCEPTIONRESNETV2, MOBILENETV3LARGE, DENSENET121, AND NASNETMOBILE

Abstract

Skin cancer cases are still increasing in the world at a very fast rate and early diagnostic intervention is pivotal in improving patient survival. While traditional clinical interpretation of skin lesion is a slow and labour-intensive subjective process which heavily relies on dermatologic expertise, this process can be greatly enhanced by automated image analysis via deep learning framework. In this study, a deep learning model was designed and tested for skin lesion classification with multi-classes. The dataset HAM10000 contains 10,015 dermoscopic images from seven different diagnostic categories. A random oversampling technique was used to overcome the basic imbalances of the datasets, and a strong image enhancement pipeline was implemented to improve the visibility of features. Five different transfer learning architectures (Xception, InceptionResNetV2, MobileNetV3Large, DenseNet121 and NASNetMobile) were used with the same custom classification head and trained under the same experimental settings. A challenging-to-classify minority sample problem was effectively solved by using a categorical focal loss. The empirical results showed that Xception architecture obtained the best performance with an F1-score of 0.98 and a rate of accuracy of 97.77%. This was followed sequentially by InceptionResNetV2 (94.40% accuracy, 0.94 F1-score), MobileNetV3Large (93.86% accuracy, 0.94 F1-score), DenseNet121 (87.35% accuracy, 0.87 F1-score), and NASNetMobile (75.12% accuracy, 0.75 F1-score). A detailed analysis was done for each model to calculate F1-scores, per-class accuracy and a confusion matrix, as well as training and validation trajectories were created. Based on the results of the experiments, the proposed pipeline is highly competitive and outperforms the recent best results obtained on the same data set. In conclusion, this research proves that transfer learning, systematic class balancing, and focal loss functions provide a computationally viable approach and highly effective method for automatic skin cancer classification, while also highlighting that the backbone architecture is the key factor that determines the effectiveness of classification under similar training conditions.