Dual hybridization method for the classification of ultrasound breast tumors
Mingue Song, Yanggon Kim · Proceedings of the 37th ACM/SIGAPP Symposium on Applied Computing · 2022
Automation of breast tumor classification has become of great importance in terms of replacing the inefficiency of manual diagnosis by individual radiologists. Most findings in literature are heavily relied on supervised learning, and such methods typically entails myriad of labelling processes. To alleviate this burden, we present an automated unsupervised learning method via double reconstruction which has the benefit of omitting data annotation step. The purpose of dual reconstruction is to compress different latent variables into a single representation in order to generate the salient attributes between lesion groups, instead of directly using each latent variable itself. Also, different types of preprocessed data are separately provided to the network as an input. Each output is expected to seize global correlation and localized variations, particularly in the tumor domain. Both representations are then hybridized as our final features to enhance the tumor discrimination. To maintain unsupervised scheme, the features are provided to the k-means clustering algorithm. The cross-validation analysis demonstrates that our method, afforded by strictly unsupervised fashion, derives significant improvements than single use of latent variables and existing pretrained supervised deep learning models Further, we affirmed the applicability of the unsupervised learning approach to this topic, and the value of sensitivity, specificity, and mean-accuracy is obtained by 93.4%, 92.26%, and 96.71%, respectively.