tsta

Home / سی و دومین کنفرانس ملی و دهمین کنفرانس بین المللی مهندسی زیست پزشکی ایران

Addressing Class Imbalance Using Difficulty-based Oversampling with Variance Control

Authors :

Zahra Asgharzadeh Bonab¹ Sina Shamekhi²

1- دانشگاه صنعتی تبریز(سهند) 2- دانشگاه صنعتی تبریز(سهند)

Keywords :

Medical image classification،oversampling،class imbalace،deep learning

Abstract :

Learning from imbalanced data is one of the difficult challenges in the field of deep learning and computer vision. Despite continuous research progress in the last few decades, learning from data with an imbalanced distribution of classes remains an important and active area of research. In real-world problems, imbalanced data often limits the applicability and practicality of deep models. The aim of this paper is to introduce a novel solution to deal with this problem in a targeted and applicable format for classification. This paper presents an innovative oversampling strategy that operates directly in feature space, rather than in pixel space. Unlike common methods such as SMOTE or ADASYN, this approach utilizes Instance Hardness level and Adversarial Neighborhood standard deviation to generate artificial instances that are aware of and fit the embedding distribution. Experimental results on three known image datasets show that this approach improves minority class recall and increases overall classification stability. The proposed approach is evaluated using deep learning architectures as embedding extractors and emphasizes the importance of different feature-embedding representations. The approach is model-independent and offers superior performance in terms of both effectiveness and computational efficiency compared to similar approaches.