Melanoma Detection with DenseNet201: A Deep Learning Approach for Early Skin Cancer Diagnosis

Melanoma is a serious and sometimes fatal disease, thus early diagnosis is crucial. In the present research work, a DenseNet201 deep learning model fine-tuned to classify skin lesions as either benign or malignant is used. The used dataset list is composed of 13,900 images of dermoscopy images that belong to the annotated dataset for feeding deep learning frameworks accordingly to the basic requirements of the platforms like TensorFlow and PyTorch by resizing the images with the 224×224 resolution of RGB format. Benign lesions have smooth margins, a uniform color and shape while malignant lesions for instance Melanoma often have irregular shape and color so as to merge seamlessly with the normal tissue. Preprocessing of the received data involved scaling where the pixel values were normalized to a range of between 0 and 1. To minimize the problem of overfitting, data was augmented through such operations as rotation, flipping and zooming. To overcome class imbalance, weighting of the loss function was used. Specificity: 0.96 for malignant lesion, sensitivity: 0.86 of the model, and F1-score of 0.90 was obtained by the model for diagnosing malignant lesion. The DenseNet201 model proved for assessing melanoma appropriately and this research showed it as a reliable technique for its early identification. Further work shall build upon the intervention using enhanced augmentation and clinical metadata integration to enhance the classification outcome.