A novel encoder–decoder wavelet model for multifocal region segmentation of TAO facial images

• Published in: Neural computing & applications 2023-09, Vol.35 (26), p.19145-19167
• Main Authors: Zhu, Haipeng, Zhou, Huifang, He, Hong, Chen, Jiayu, Song, Xuefei, Li, Kunhao, Zhou, Lei
• Format: Article
• Language: English
• Subjects: Artificial Intelligence; Autoimmune diseases; Coders; Computational Biology/Bioinformatics; Computational Science and Engineering; Computer Science; Data Mining and Knowledge Discovery; Diagnosis; Discrete Wavelet Transform; Feature extraction; Formability; Image Processing and Computer Vision; Image segmentation; Medical imaging; Original Article; Probability and Statistics in Computer Science; Signs and symptoms; Wavelet transforms
• ISSN: 0941-0643; 1433-3058
• DOI: 10.1007/s00521-023-08727-2

Thyroid-associated ophthalmopathy (TAO) is an organ-specific autoimmune disease that has a significant impact on patients` life and health of. Clinically, the Clinical Activity Score (CAS) is one of the crucial methods for the early diagnosis of TAO. However, due to the diversity of TAO symptoms, utilizing medical expertise to artificially obtain CAS scores is challenging and highly dependent on personal subjectivity. Therefore, accurate identification of TAO regions segmented by scientific techniques is one of the essential prerequisites for the objective acquisition of the CAS scores. In this study, an encoder–decoder wavelet model (EDWM) with multiple-scale cascaded attention mechanism (MCAM) and residual deformable convolution (RDC) was proposed for multifocal region segmentation of TAO from facial images. The proposed method employs the discrete wavelet transform (DWT) to construct an encoder structure for the coarse feature extraction of the diseased regions. The inverse wavelet transform (IWT) is designed to build a decoder structure for resolution recovery. Meanwhile, the MCAM is developed to extract finer features of adjacent wavelet scales in the encoder structure by suppressing the background and focusing on the coarse segmentation of the diseased regions. The RDC is ultimately utilized for enlargement of arbitrary receptive fields and the accurate multi-segmentation task in different regions. In comparison with other selected benchmark models, the EDWM has, respectively, achieved state-of-the-art segmentation performance with 93.12% and 0.804 of the precision and the MIoU when tested on the images of 600 TAO patients. Since the EDWM is characterized by compact structure, interpretability, and strong feature extraction capability, it can provide a much more reliable and scientific basis for the early detection and diagnosis of TAO, reducing reliance on subjective experience in obtaining CAS scores.