Exploring Explainable Melanoma Classification: Leveraging Pre-trained Deep Learning Model on MED-NODE Dataset

Skin cancer, an extremely common and potentially fatal condition, emphasizes the critical importance of timely and precise detection. This study presents a thorough examination of dermatological image classification using deep learning models on the Med Node dataset. Five prominent models, including InceptionV3, Xception, VGG19, EfficientNetB1, and DenseNet201, were assessed for their ability to discern between melanoma and naevus instances. Noteworthy variations in performance metrics were observed, with Xception standing out with an exceptional accuracy of 95.88% and perfect precision and recall for both classes. In contrast, InceptionV3 demonstrated a balanced trade-off between precision and recall. VGG19 exhibited comparatively lower performance, while EfficientNetB1 and DenseNet201 showcased outstanding accuracy, with DenseNet201 leading with a remarkable 96.47%. A subsequent statistical analysis using z-scores and two-tailed p-values confirmed the significant differences among the top three models (EfficientNetB1, Xception, and DenseNet201). The study compared the proposed DenseNet201 model with the existing PECK Ensemble model. The results indicated a substantial 5% performance improvement of We have also added explainable AI (XAI) using Lime to visualize lesion section. Z-score of models is also calculated to check its reliability.