Robustness of Image Classification on Imbalanced Datasets Using Capsules Networks

The purpose of this study is to assess the robustness of the capsule networks for the image classification tasks on imbalanced dataset. In fact, in many real-life situations, the distribution of samples is skewed, with representatives of certain classes appearing much more frequently. This raises a problem for learning algorithms, as they will be biased in favor of the majority group. At the same time, the minority class is usually the most important, because, despite its rarity, it may carry important and useful information. In this work, we firstly design a capsule networks architecture that we apply to a medical dataset extremely imbalanced provided by the Broad Institute, then we randomly subsampled that dataset and we compare the result to CNN architectures. To gauge our model’s ability to perform well on diverse datasets, we tested it on a specifically configured version of the CIFAR10 dataset. This version has a class imbalance ratio of 0.1, meaning some classes have significantly fewer data points than others. This mimics the real-world scenario of imbalanced datasets. We then compared our model’s performance on this dataset with the findings of recent research in the field. To evaluate results, we use the following metrics: Accuracy, Precision, Recall, and F1-score. The results of experiments show that capsules networks can give good results on highly imbalanced datasets (BBBC41 and BBBC042) and with small observations and can be in certain instance better than convolutional neural networks. While our Capsule network architecture achieved a recall of 0.7 on the imbalanced CIFAR10 dataset, recent studies utilizing focal loss and complement cross-entropy with convolutional neural networks (CNNs) have demonstrated superior performance. These findings highlight the impact of both model architecture and data characteristics. Our model achieved better results on datasets with more severe imbalances (1:700 and 1:15) compared to CIFAR10’s milder imbalance (1:10). This underscores the importance of considering both model and dataset factors when evaluating performance on imbalanced data. In conclusion, Capsule networks show promise for imbalanced learning, but they are not a universal solution. Further research is needed to explore their effectiveness across a wider range of imbalanced data scenario.