Instance Segmentation Based on Deep Convolutional Neural Networks and Transfer Learning for Unconstrained Psoriasis Skin Images

• DOI: 10.3390/app11073155
• Publication Date: 2021-04-01T14:44:01Z
• Authors: Guo-Shiang Lin; Kuan-Ting Lai; Jian-Ming Syu; Jen-Yung Lin; Sin-Kuo Chai
• Keywords: Minimum bounding box; Feature (linguistics); Pyramid (geometry); Transfer of learning

In this paper, an efficient instance segmentation scheme based on deep convolutional neural networks is proposed to deal with unconstrained psoriasis images for computer-aided diagnosis. To achieve segmentation, the You Only Look At CoefficienTs (YOLACT) network composed of backbone, feature pyramid (FPN), Protonet, and prediction head used images. The backbone extract maps from image, FPN designed generate multiscale effectively classifying localizing objects multiple sizes. predict classification information, bounding box mask coefficients objects. Some prototypes generated by Protonet are combined estimate pixel-level shapes images, YOLACT++ a pretrained model retrained via transfer learning. evaluate performance scheme, different severity levels collected testing. As subjective testing, regions normal skin areas can be located classified well. four indices were higher than 93% after cross validation. About object localization, Mean Average Precision (mAP) rates at least 85.9% efficiency, frames per second (FPS) rate reached up 15. addition, F1_score execution speed those Mask Region-Based Convolutional Neural Networks (R-CNN)-based method. These results show that not only detect but also distinguish pixels background Furthermore, outperforms R-CNN-based method