Guided Random Mask: Adaptively Regularizing Deep Neural Networks for Medical Image Analysis by Potential Lesions

Abstract

Data augmentation is a critical regularization method that contributes to numerous state-of-the-art results achieved by deep neural networks (DNNs). The visual interpretation method demonstrates that the DNNs behave like object detectors, focusing on the discriminative regions in the input image. Many studies have also discovered that the DNNs correctly identify the lesions in the input, which has been confirmed in the current work. However, for medical images containing complicated lesions, we observe the DNNs focus on the most prominent abnormalities, neglecting sub-clinical characteristics that may also help diagnosis. We speculate this bias may hamper the generalization ability of DNNs, potentially causing false predicted results. Based on this consideration, a simple yet effective data augmentation method called guided random mask (GRM) is proposed to discover the lesions with different characteristics. Visual interpretation of the inference result is used as guidance to generate random-sized masks, forcing the DNNs to learn both the prominent and subtle lesions. One notable difference between GRM and conventional data augmentation methods is the association with the training phase of DNNs. The parameters in vanilla augmentation methods are independent of the training phase, which may limit their effectiveness when the scale and appearance of region-of-interests vary. Nevertheless, the effectiveness of the proposed GRM method evolves with the training of DNNs, adaptively regularizing the DNNs to alleviate the over-fitting problem. Moreover, the GRM is a parameter-free augmentation method that can be incorporated into DNNs without modifying the architecture. The GRM is empirically verified on multiple datasets with different modalities, including optical coherence tomography, x-ray, and color fundus images. Quantitative experimental results show that the proposed GRM method achieves higher classification accuracy than the commonly used augmentation methods in multiple networks. Visualization analysis also demonstrates that the GRM can better localize lesions than the vanilla network.

Keywords

• deep neural networks
• data augmentation
• regularization
• medical image analysis