DBTN: An adaptive neural network for multiple-disease detection via imbalanced medical images distribution

Acquiring medical images while maintaining patient information confidentiality is a difficult task, which leads to a lack of sufficient data for deep learning-based disease detection. To address the challenges of few-shot learning in such scenarios, researchers generally resort to the transfer learning ability of the pre-train model. However, in practice, the pre-trained model only updates its fully connected layers or a few other layers with limited data examples while holding its many layers unchanged. Nonetheless, such transfer learning paradigms are not expected to alter the layer structure of the model. Consequently, the benefits of pre-trained models are limited, especially when the target dataset deviates from the source dataset in terms of data pattern. Moreover, the time cost of transfer retraining should also be considered. To alleviate these impacts, we propose a novel paradigm for multiple disease detection via medical images that combined a meta-heuristic algorithm and transfer learning. We also present a novel optimization algorithm and design two novel adaptive frameworks (BTL, DTL), and four adaptive neural networks (BVTN, BSTN, DVTN, and DSTN). In our paradigm, pre-trained models could adaptively adjust the feature extraction ability by reshaping a few network layers to accommodate the data distribution differences between source and target datasets. Thus, the negative impact of such differences can be effectively alleviated. We conduct experiments on three datasets (brain magnetic resonance, chest X-ray, and skin image datasets) with different data distributions. Experimental results show that our paradigm can effectively mitigate the impact of differences in data distribution, and that our method outperforms four state-of-the-art methods on all three datasets.