Unsupervised learning method via triple reconstruction for the classification of ultrasound breast lesions

Automatic classification of ultrasound breast tumors has been pragmatically alleviating several defects raised in manual diagnosis and utilized as a second opinion to help diagnostic procedures. The extensive works in literature have excessively tackled this task in a supervised learning frame, but the need for large amounts of annotated data followed by such methods represents one of the fatal drawbacks. To address this, this work proposes an unsupervised learning method based on encoder-decoder scheme that allows to avoid the demand of labeling information in the entire discrimination step.A customized convolutional autoencoder is devised as a backbone architecture. The first step of the scenario is to pre-train the network in a generic way. Then, our strategy is to arbitrarily change the target mapping of the object in the reconstruction loss in order to encapsulate different details in the embedding space. Also, the sequential form of autoencoder is introduced to perform gradual reconstruction, and the mutual information derived from three different representations was generated. The discrimination was performed via subsequent clustering, and thus the aim of this study is to suggest the potentiality of unsupervised learning framework in this field.To verify the effectiveness, experiments were conducted on two publicly available datasets, BUSI and UDIAT. The comprehensive results demonstrate that the mutual information is capable of remarkably enhancing discriminant capability and outperforms the actual supervised learning-based state-of-the-art methods in terms of discrimination. Further, our method has advantages of not requiring any region of interest extraction, localization, or annotated information during training.

Automated summary

This study proposes an unsupervised learning method based on an encoder-decoder scheme for automatic classification of ultrasound breast tumors, demonstrating its potential in enhancing discriminant capability. By introducing the sequential form of autoencoder and changing the target mapping of the object, the study successfully achieved effective classification without the need for annotated information.