Skin cancer detection using harmonic brown bear optimization enabled transfer learning

Skin cancer is a potentially fatal disease that can be successfully treated if identified at an early stage. However, early diagnosis is often difficult due to the visual similarity of malignant and benign skin lesions. To address this challenge, this paper proposes the Convolutional Neural Network-based Transfer Learning enhanced by Harmonic Brown Bear Optimization (CNN-TL_Hr-BOA) to improve the accuracy and robustness of skin cancer detection. The process begins with image denoising using the Medav filter to eliminate artifacts and noise. Next, skin lesion segmentation is performed using the Position and Context Information Fusion Network (PCF-Net), which effectively isolates lesion regions by combining positional and contextual cues. The segmented images are further improved through superpixel-mixing-based image augmentation, which diversifies the training data by preserving important structural details. Subsequently, image-level features and Reductant Discrete Wavelet Transform (RDWT) are extracted. These features are then fed into a CNN model initialized with pre-trained DenseNet weights under a transfer learning setup. To enhance model performance, Hr-BOA, a hybrid metaheuristic that combines Harmonic Analysis with the Brown Bear Optimization Algorithm, is introduced to fine-tune the CNN's hyperparameters efficiently. The proposed CNN-TL_Hr-BOA framework is evaluated using the SIIM-ISIC Melanoma Classification dataset, where it demonstrates superior detection capability. The model achieves an accuracy of 91.754 %, a True Positive Rate (TPR) of 93.755 %, and a True Negative Rate (TNR) of 89.766 % using 90 % of the dataset for training. These results confirm the effectiveness of the proposed approach in accurately identifying skin cancer at early stages.