Abstract
Data augmentation is an effective method to prevent model overfitting in deep learning, especially in medical image classification where data samples are small and difficult to obtain. In recent years, different data augmentation methods, such as those based on single data transformation, multiple data mixing, and learning data distribution, have been proposed one after another, but there has never been a systematic system to evaluate various data augmentation methods. An impartial and comprehensive data augmentation evaluation system not only can assess the benefits and drawbacks of existing augmentation approaches in a specific medical image classification but also can provide an effective research direction for the subsequent proposal of new medical image data augmentation methods, thereby advancing the development of auxiliary diagnosis technology based on medical images. Therefore, this paper proposes an objective and universal evaluation system for different data augmentation methods. In this method, different augmented methods are evaluated objectively and comprehensively in terms of classification accuracy and data diversity by using existing large public data sets. The method is universal and easy to operate. To imitate the prevalent small-sized data sets in deep learning, an equal-interval sampling technique based on similarity ranking is presented to select samples from large public data sets and construct a subset that can fully reflect the original set. The augmented data sets are then created using various data augmentation approaches based on the small-sized data sets. Finally, different data augmentation strategies are objectively and fully evaluated based on the comprehensive scores of classification accuracy and data diversity following data augmentation. The validity and feasibility of the suggested sampling method and assessment system in this study are demonstrated by experimental findings on numerous data sets.
Highlights
Data augmentation is an effective method to prevent model overfitting in deep learning, especially in medical image classification where data samples are small and difficult to obtain
In order to objectively and comprehensively evaluate the performance of different augmentation methods, this paper considers the data set after augmentation from three dimensions
In addition to the two medical imaging data sets, we selected two additional data sets from other fields to verify the broad applicability of our equal-interval sampling algorithm. ese data sets are: medical CT image data set (DeepLesion) [19], Pneumonia X-ray image data set [20], ImageNet [21], and CIFAR10 [22], and use six different sampling methods to sample the above two data sets, so as to generate small-sized data sets. e classification accuracy difference between the generated small-sized data set and the original data set, as well as the Frechet Inception Distance (FID) between them, were used as evaluation indexes to evaluate the performance of different sampling methods
Summary
To prepare data for the suggested data augmentation method assessment system, a small-sized data set should be taken from the original large data set, which adheres and corresponds to the original data set’s distribution and can completely represent it. According to the previously selected sampling number, equal-interval sampling is performed for this category set, and the Frechet Inception Distance (FID) index [14] between the sampled data set and the original complete data set is applied as a secondary verification. To construct a small-sized data set with similar distribution to the original large-scale data set, which is prepared for subsequent evaluation of various data augmentation methods. E Frechet Inception Distance, or FID, is a metric for assessing the quality of the produced images that were designed to calculate the performance of generative adversarial networks To construct a small-sized data set with similar distribution to the original large-scale data set, which is prepared for subsequent evaluation of various data augmentation methods. e Frechet Inception Distance, or FID, is a metric for assessing the quality of the produced images that were designed to calculate the performance of generative adversarial networks
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
