Abstract: The skin is the body's largest organ. Melanoma is a highly malignant skin tumor for which there is currently no effective treatment. Therefore, the best approach is to perform early diagnosis and treatment in initial stages. Thus, with an attitude of early detection and treatment to increase post-operative survival time, melanoma detection in its initial stages becomes especially important. In the era of advanced deep learning (DL), combining DL with medical imaging for computer-aided diagnosis (CAD) and lesion segmentation can not only increase the efficiency of doctors’ work but also make it more convenient for the general public to pay attention to and monitor their own health. In this paper, a novel Deep Feature Pyramid Network (DFP-Net) DL model is proposed. It uses a Feature Pyramid Network (FPN) as a basis, with ResNet-101 as the backbone network, and incorporates newly designed Residual Receptive Field Blocks (RRFBs), to increase the receptive field while also enhancing its feature extraction capabilities. A novel Squeeze-and-Excitation Atrous Spatial Pyramid Pooling (SEASPP) module is added to capture and filter contextual information better after the feature extraction process. Newly proposed Depthwise Over-parameterized Convolutional Block Attention Modules (DOCBAMs) are introduced after each top-down step to further refine the features, which are then fused to produce the final output. Experiments, conducted on two datasets, demonstrate the superiority of the proposed DFP-Net model compared to state-of-the-art models, based on multiple metrics.

Keywords: melanoma, medical imaging, segmentation, computer-aided diagnosis (CAD), deep learning (DL), Feature Pyramid Network (FPN), Deep Feature Pyramid Network (DFP-Net)

Pages: 387-402