Pixels to Classes: Intelligent Learning Framework for Multiclass Skin Lesion Localization and Classification. (March 2021)

• Record Type: Journal Article
• Authors: Khan, Muhammad Attique; Zhang, Yu-Dong; Sharif, Muhammad; Akram, Tallha
• Is Part Of: Computers & electrical engineering. Volume 90(2021)
• Publication Date: 2021-03
• Subjects: Skin Cancer -- Lesion Segmentation -- CNN Architecture -- Recognition; Computer engineering -- Periodicals; Electrical engineering -- Periodicals; Electrical engineering -- Data processing -- Periodicals; Ordinateurs -- Conception et construction -- Périodiques; Électrotechnique -- Périodiques; Électrotechnique -- Informatique -- Périodiques; Computer engineering; Electrical engineering; Electrical engineering -- Data processing; Periodicals; Electronic journals; 621.302854
• DOI: 10.1016/j.compeleceng.2020.106956
• Languages: English
• ISSNs: 0045-7906

Abstract

Highlights: Modified Mask RCNN based skin lesion segmentation is performed. 24-Layered CNN architecture is implemented. Fully Connected layer is utilized for features mapping Softmax classifier is employed for lesion type recognition Abstract: A novel deep learning framework is proposed for lesion segmentation and classification. The proposed technique incorporates two primary phases. For lesion segmentation, Mask recurrent convolutional neural network (MASK R-CNN) based architecture is implemented. In this architecture, Resnet50 along with feature pyramid network (FPN) is utilized as a backbone. Later, fully connected layer-based features are mapped for the final mask generation. In the classification phase, 24-layered convolutional neural network architecture is designed, which performs activation based on the visualization of higher features. Finally, best CNN features are provided to softmax classifiers for final classification. Three datasets (i.e. PH2, ISBI2016, and ISIC2017) are utilized for the validation of the segmentation process, whilst HAM10000 dataset is utilized for the classification. From the results, it is concluded that the proposed method outperforms several existing techniques, based on the selected set of parameters including sensitivity (85.57%), precision (87.01%), F1- Score (86.28%), and accuracy (86.5%).