Decision Boundary Re-Sampling in Imbalanced Learning for Ulcer Detection

Abstract

Data imbalance problem between normal and lesion endoscopy images makes it difficult to employ deep learning approaches in automatic Ulcer detection and classification. Due to the large variety of normal images in their appearance, characterizing ulcer with limited training samples is not a trivial task. In this work, we propose decision boundary re-sampling (DBR) in imbalanced learning that extrapolates ulcer samples in a latent space of deep convolutional neural network. Proposed method shows improved ulcer classification performance on wireless endoscopy images compared to state-of-the-art methods.

Class specific biased extrapolation of images in latent space for imbalanced image classification

“…We employ decision boundary re-sampling [8] and biased extrapolation [15] schemes in our study and evaluation. Biased extrapolation between two classes can reveal the characteristics of the classes such as diversity and complexity.…”

“…Biased extrapolation re-sampling is performed in each minor-major instance pair. Decision boundary re-sampling [8] adds the new sample on the minor-major class boundary. On the other hand, biased extrapolation finds the optimal location of new samples based on diverse bias condition.…”

“…Figure 1 shows precision-recall curves of all minor-major pairs before and after biased extrapolation. Each graph compares precisionrecall curves of five cases: No extrapolation, Extrapolation with 30% randomly chosen normal samples (Random) [8], Extrapolation with bottom 30% of normal sample probability (Bottom), Extrapolation with top 30% of normal sample probability (Top), Asymmetry extrapolation based on the equation (1).…”

“…One of the classes with complex truth distribution with within-class diversity may require more training samples outside of the current training sample distribution. Sample extrapolation methods [15,6,7,8,9,13] are the solution for the problem. Han et al [6] create new minor samples on the borderline with other classes rather than within minor class samples.…”

“…Han et al [6] create new minor samples on the borderline with other classes rather than within minor class samples. Lee et al [7,8] propose a feature space extrapolation with deep convolutional neural networks. Decision boundary re-sampling (DBR) finds new minor samples on the decision boundary of latent space.…”