Semi-Supervised Medical Image Classification with Neighborhood Adaptive Attention

Semi-supervised learning utilize unlabeled data to improve model performance, which is of great significance in medical imaging tasks. However, due to the high inter class similarity and large intra class variance in medical data, this still poses challenges. To address these challenges, a neighborhood adaptive attention (NAA) mechanism is introduced, which extends to surrounding attributes based on mutual correlation, considers the lesion area and its highly correlated pathological attributes, enabling the model to learn additional semantic information and improve diagnostic performance. Unlike previous semi-supervised frameworks, sample information below the threshold is not discarded. In contrast, class imbalance comparison loss (CICL) is constructed by combining the information of samples above and below the threshold. Our method has been evaluated in challenging medical image classification tasks: classifying lesions from ultrasound images and classifying multiple ophthalmic diseases from optical coherence tomography images of the retina. The results show that this method outperforms supervised baselines and other state-of-the-art methods in all tasks.