A new method proposed to Melanoma-skin cancer lesion detection and segmentation based on hybrid convolutional neural network

The number of deaths due to melanoma skin cancer has rapidly increased in recent years. The timely diagnosis of the lesions of melanoma skin cancer can potentially increase the survival rate of such a chronic disease. However, the detection of these lesions is a challenging task, especially in the presence of occlusions, such as clinical artifacts, blood vessels, and color contrast variation, etc. The current state-of-the-art detection and segmentation methods are based on fully convolutional neural networks, which utilize an encoder-decoder method. However, these methods produce coarse segmentation masks due to the loss of location information during the encoding layers. To overcome these challenges, this study proposes a highly effective Hybrid detection and segmentation method based on the integration of RetinaNet and MaskRCNN, which utilizes a pyramid module of lateral connections and top-down paths to compensate for the loss of spatial features information. The proposed method is trained and validated on Melanoma-ISIC-2018 and PH2 datasets. Experiment results on the unseen PH2 dataset illustrate the improved generalization ability of the method. The efficacy with other methods such as Encoder-Decoder, Generative Adversarial Network(GAN), UNet Deep Convolutional Neural Network-support vector machine(DCNN-SVM), Encoder-Fully Connected Network(EFCN), Enhanced Convolutional-Deconvolutional Networks(ECDNs), UNet, and Handcrafted has also been compared. The results, show that the proposed method outperforms above methods by 7.7%, 12.9%, 11.4%, 14.4%, 14.9%, 18.6%, 25.1%, respectively, in terms of accuracy. It is envisaged that with reliable accuracy, this method can be introduced for clinical practices in the future.