Imbalance multiclass problem: a robust feature enhancement-based framework for liver lesion classification

The classification of liver lesions in CT images is essential for the diagnosis and treatment of liver diseases. Since the characteristics of different classes of lesions are similar and the degree of differentiation is not obvious, it is difficult to accurately classify different classes of liver lesions, especially imbalanced multiclass distribution. To this end, we propose a novel feature enhancement-based framework for imbalanced liver lesion classification. Specifically, a liver lesion processing method is introduced to expand data based on Augmentor. Besides, based on Augmentor, we design dual feature enhancement, which is based on feature refinement extraction and feature global correlation, to enhance feature reconstruction ability. To further alleviate the problem of class imbalance, the improved loss function based on standard cross-entropy (CE) is also adopted to make the network pay more attention to the class with fewer samples. Experiments on 551 liver lesions in 120 patients showed that: (1) This framework improved the classification performance of liver tumors under imbalanced multiclass data distribution; (2) The dual feature enhancement was lightweight and efficient which enhanced the semantic expression of overall features without introducing additional parameters.