Unsupervised medical image feature learning by using de-melting reduction auto-encoder

Unsupervised feature learning is a fundamental and highly prioritized problem in medical image analysis. Although it has shown considerable improvements, it remains challenging because of its weak feature expression ability, low model-learning efficiency, and weak robustness. To address these limitations, a novel unsupervised feature learning method in the medical image classification task, named de-melting reduction auto-encoder (DMRAE), is proposed in this study. A joint fusion network structure is constructed; it not only improves the expression of target features but also reduces the loss of feature decoding and parameters. To obtain a robust solution, a newly designed decomposed-reconstructed loss function is used to strengthen the semantic context between adjacent feature extractor layers, successfully avoiding the insufficient model-learning ability from the single optimization objective and improving the quality of the extracted features. Finally, extensive experiments on datasets consisting of 400 breast ultrasonographic images and 6000 lung computed tomography images are conducted to demonstrate the effectiveness of the proposed method. Experimental results reveal that the DMRAE significantly reduces the annotation effort and outperforms existing methods by a significant margin.