MMTDNN: Multi-View Massive Training Deep Neural Network for Segmentation and Detection of Abnormal Tissues in Medical Images

Abstract

Purpose: Automated segmentation of abnormal tissues in medical images is considered as an essential part of those computer-aided detection and diagnosis systems which analyze medical images. However, automated segmentation of abnormalities is a challenging task due to the limitations of imaging technologies and complex structure of abnormalities, including low contrast between normal and abnormal tissues, shape diversity, appearance inhomogeneity, and the vague boundaries of abnormalities. Therefore, more intelligent segmentation techniques are required to tackle these challenges. Materials and Methods: In this study, a method, which is called MMTDNN, is proposed to segment and detect medical image abnormalities. MMTDNN, as a multi-view learning machine, utilizes convolutional neural networks in a massive training strategy. Moreover, the proposed method has four phases of preprocessing, view generation, pixel-level segmentation, and post-processing. The International Symposium on Biomedical Imaging (ISBI)-2016 dataset is used for the evaluation of the proposed method. Results: The performance of the proposed method has been evaluated on the task of skin lesion segmentation as one of the challenging applications of abnormal tissue segmentation. Both qualitative and quantitative results demonstrate outstanding performance. Meanwhile, the accuracy of 0.973, the Jaccard index of 0.876, and the Dice similarity coefficient of 0.931 have been achieved. Conclusion: In conclusion, the experimental result demonstrates that the proposed method outperforms state-of-the-art methods of skin lesion segmentation.