A semi-supervised resampling method for class-imbalanced learning

Jiang, Zhen; Zhao, Lingyun; Lu, Yu; Zhan, Yongzhao; Mao, Qirong

doi:10.1016/j.eswa.2023.119733

Cited by 23 publications

(9 citation statements)

References 47 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Analysis of Oversampling and under-sampling data distribution was performed by [44] [49] for handling imbalanced data problems in classification. The proposed transfer learning model includes three modules, i.e., 1) Active sampling module, 2) Real-time data augmentation module and 3) DenseNet module.…”

Section: Related Workmentioning

confidence: 99%

A Unified Approach Addressing Class Imbalance in Magnetic Resonance Image for Deep Learning Models

Cui,

Li,

Yang

et al. 2024

IEEE Access

View full text Add to dashboard Cite

Medical image datasets, particularly those comprising Magnetic Resonance (MR) images, are essential for accurate diagnosis and treatment planning. However, these datasets often suffer from class imbalance, where certain classes of abnormalities have unequal representation. Models trained on imbalanced datasets can be biased towards the prominent class, leading to misclassification. Addressing class imbalance problems is crucial to developing robust deep-learning MR image analysis models. This research focuses on the class imbalance problem in MR image datasets and proposes a novel approach to enhance deep learning models. We have introduced a unified approach equipped with a selective attention mechanism, unified loss function, and progressive resizing. The selective attention strategy identifies prominent regions within the underlying image to find the feature maps, retaining only the relevant activations of the minority class. Finetuning of the multiple hyperparameters was achieved using a novel unified loss function that plays a vital role in enhancing the overwhelming error performance for minority classes and accuracy for common classes. To address the class imbalances phenomenon, we incorporate progressive resizing that can dynamically adjust the input image size as the model trains. This dynamic nature helps handle class imbalances and improve overall performance. The research evaluates the proposed approach's effectiveness by embedding it into five state-of-the-art CNN models: UNet, FCN, RCNN, SegNet, and Deeplab-V3. For experimental purposes, we have selected five diverse MR image datasets, BUS2017, MICCAI 2015 head and neck, ATLAS, BRATS 2015, and Digital Database Thyroid Image (DDTI), to evaluate the performance of the proposed approach against state-of-the-art techniques. The assessment of the proposed approach reveals improved performance across all metrics for five different MR imaging datasets. DeepLab-V3 demonstrated the best performance, achieving IoU, DSC, Precision, and Recall scores of 0.893, 0.953, 0.943, and 0.944, respectively, on the BUS dataset. These scores indicate an improvement of 5% in DSC, 6% in IoU, 4% in precision, and approximately 4% in recall compared to the baseline. The most significant increases were observed in the ATLAS and LiTS MICCAI 2017 datasets, with a 5% and 7% increase in IoU and DSC over the baseline (DSC = 0.628, DSC = 0.695) for the ATLAS dataset, and a 5% and 9% increase in IoU and DSC for the LiTS MICCAI 2017 dataset.

show abstract

Section: Related Workmentioning

confidence: 99%

A Unified Approach Addressing Class Imbalance in Magnetic Resonance Image for Deep Learning Models

Cui,

Li,

Yang

et al. 2024

IEEE Access

View full text Add to dashboard Cite

show abstract

“…Note that the removal of majority samples may lead to the loss of important information. Correspondingly, several approaches were proposed to address this problem by utilizing the data distribution to improve the undersampling (Jiang et al, 2023; Peng et al, 2019). Recently, some reported works demonstrated the effectiveness of Deep Neural Network (DNN) on class‐imbalanced datasets.…”

Section: Related Workmentioning

confidence: 99%

A boosted co‐training method for class‐imbalanced learning

2023

Self Cite

View full text Add to dashboard Cite

Class imbalance learning (CIL) has become one of the most challenging research topics. In this article, we propose a Boosted co-training method to modify the class distribution so that traditional classifiers can be readily adapted to imbalanced datasets. This article is among the first to utilize pseudo-labelled data of co-training to enlarge the training set of minority classes. Compared with existing oversampling methods which generate minority samples based on labelled data, the proposed method has the ability to learn from unlabelled data and then decrease the risk of overfitting. Furthermore, we propose a boosting-style technique which implicitly modifies the class distribution and combines it with co-training to alleviate the bias towards majority classes. Finally, we collect two series of classifiers generated during Boosted co-training to build an ensemble for the classification. It further improves the CIL performance by leveraging the strength of ensemble learning. By taking advantage of the diversity of co-training, we also contribute a new approach to generating base classifiers for ensemble learning. The proposed method is compared with eight state-of-the-art CIL methods on a variety of benchmark datasets. Measured by G-Mean, F-Measure, and AUC, Boosted co-training achieves the best performances and average ranks on 18 benchmark datasets. The experimental results demonstrate the significant superiority of Boosted co-training over other CIL methods. K E Y W O R D S boosting, class-imbalanced learning, co-training, over-sampling, pseudo-labelled data | INTRODUCTIONMany real-world classification tasks suffer from the class imbalance problem, where minority classes are highly underrepresented as compared to majority classes. Note that traditional classifiers are designed to output the hypothesis that minimizes the overall prediction error. As a result, they are apt to be biased towards majority classes and thereby perform poorly on minority classes (Kaur et al., 2019). However, minority classes are usually more valuable in real applications, such as fraud detection, medical diagnosis, spam classification, and many others. For example, in rare disease diagnoses, a classifier that identifies all patients as normal cases is useless even if it achieves 99% accuracy. Therefore, learning from class-imbalanced data has become one of the most challenging topics in machine learning. Numerous CIL techniques have been proposed over the past decades, which can be roughly grouped into the following two categories:i. Data-level methods try to preprocess (e.g., oversampling (Chawla et al., 2002) or undersampling (Kubat & Matwin, 1997)) a dataset to make it suitable for standard classification algorithms. This approach is classifier-independent. However, generating a perfectly balanced distribution does not always provide an optimal result for classification tasks (Wu & Chang, 2003).

show abstract

“…Cluster-based sampling methods are commonly used to uncover the underlying data structure in resampling techniques (Jiang, Zhao, et al, 2023) and are more effective than random or synthetic sampling for identifying 𝑃 𝑖 ∈ . Several approaches have been proposed to address CIL, for instance, Jo and Japkowicz (2004) applied k-means clustering beforehand, assuming clusters  and  as disjunct subsets.…”

Section: Cluster-based Samplingmentioning

confidence: 99%

“…One approach that has attracted significant interest in recent years is cluster analysis, which is well-suited for identifying data distributions for resampling purposes (Jiang, Zhao, et al, 2023, Lu et al, 2022. In these methods, majority and minority samples are assigned to distinct clusters, and either oversampling or undersampling is performed within these clusters to achieve a balanced class distribution.…”

Section: Introductionmentioning

confidence: 99%