A Deeply Connected Model for High-Precision Classification of Benign and Malignant Skin Lesions

This work presents DenseNet-SLC, a new high performance model for the binary classification of skins lesion as benign or malignant based on dermatoscopic images. Taking the DenseNet121 design as the base, the model adds the first 121 layers with the ImageNet weights updated and adds a few more layers, including MaxPooling, BatchNormalization, and Dropout (Drop off rates of 30% and 50%) to empower the model and reduce overfitting. Our DenseNet-SLC model had a training accuracy of 96% and a validation accuracy 90% with both the accuracy and loss showing no fluctuations after initial oscillations lending to generalization. These outcomes are higher than the current models, which offer 85%-90% power of reflection, at most. Due to the high accuracy and stability of DenseNet-SLC, information derived from it can help dermatologists facilitate the prompt diagnosis of skin carcinoma that will help in improving the prognosis of a patient's disease. The disclosed ability to distinguish between malignant and benign lesions quickly and accurately situates it as a valuable tool in practical diagnostic applications for clinical practice.