<p>This work proposes a skin lesion classification approach that uses an adjacent dynamic and fixed convolutional kernel-based convolutional neural network (DyFi-CNN) to extract fine texture features from the lesion region. This work initially pre-process the image using the median filter followed by a contrast enhancement process. The lesion region is separated from its background using Otsu’s segmentation approach. Further, the boundary pixels of the lesion are enhanced using the boundary enhancement process, since the boundary of the lesion region plays a major role in classifying the lesion types. Finally, the proposed DyFi-CNN is utilized to extract the fine texture descriptors from the RGB channels and classify the skin lesions. The proposed DyFi-CNN has two different types of kernels namely the dynamic kernels and the fixed kernels. The dynamic kernels are placed between two adjacent fixed kernels. The coefficients of the dynamic kernels are updated throughout the convolution process, while the coefficients of the fixed kernels remain constant throughout the convolution process. The evaluation of the approach was done with the dermoscopic images collected from the HAM-10&#xa0;K dataset using different classification metrics. The HAM-10&#xa0;K dataset results in F1-score, sensitivity, accuracy, and specificity of 94.67%, 95.17%, 98.02%, and 93.83% respectively in classifying 7 different lesion types.</p>