[Paper Review] General Purpose (GenP) Bioimage Ensemble of Handcrafted and Learned Features with Data Augmentation

Loris Nanni, Sheryl Brahnam+2 more|arXiv (Cornell University)|Apr 17, 2019

Image Processing Techniques and Applications80 references17 citations

TL;DR

This paper proposes a General Purpose (GenP) bioimage ensemble that combines handcrafted features, deep learning via CNNs, and data augmentation using PCA/DCT-based methods to achieve state-of-the-art performance in bioimage classification. By integrating diverse descriptors with an SVM ensemble and CNN ensemble, and applying novel augmentation techniques, the method achieves high accuracy without dataset-specific hyperparameter tuning, minimizing overfitting risks.

ABSTRACT

Bioimage classification plays a crucial role in many biological problems. In this work, we present a new General Purpose (GenP) ensemble that boosts performance by combining local features, dense sampling features, and deep learning approaches. First, we introduce three new methods for data augmentation based on PCA/DCT; second, we show that different data augmentation approaches can boost the performance of an ensemble of CNNs; and, finally, we propose a set of handcrafted/learned descriptors that are highly generalizable. Each handcrafted descriptor is used to train a different Support Vector Machine (SVM), and the different SVMs are combined with the ensemble of CNNs. Our method is evaluated on a diverse set of bioimage classification problems. Results demonstrate that the proposed GenP bioimage ensemble obtains state-of-the-art performance without any ad-hoc dataset tuning of parameters (thus avoiding the risk of overfitting/overtraining).

Motivation & Objective

• To develop a general-purpose framework for bioimage classification that avoids dataset-specific hyperparameter tuning.
• To improve classification performance by combining handcrafted features with deep learning models.
• To introduce novel data augmentation techniques based on PCA and DCT to enhance model robustness.
• To create a highly generalizable feature descriptor set applicable across diverse biological image datasets.
• To demonstrate state-of-the-art performance across multiple bioimage benchmarks using an ensemble of SVMs and CNNs.

Proposed method

• Proposes three new data augmentation methods based on PCA and DCT to increase training data diversity.
• Employs local and dense sampling features to extract rich spatial patterns from bioimages.
• Trains separate Support Vector Machines (SVMs) on individual handcrafted descriptors for robust classification.
• Combines multiple pre-trained CNNs into an ensemble to improve generalization and prediction accuracy.
• Integrates handcrafted features and deep features via early or late fusion strategies within the ensemble framework.
• Applies the ensemble approach across multiple bioimage datasets without any dataset-specific hyperparameter tuning.

Experimental results

Research questions

• RQ1 Can a unified ensemble of handcrafted and learned features improve bioimage classification performance across diverse datasets?
• RQ2 How do PCA- and DCT-based data augmentation techniques enhance the generalization of bioimage classifiers?
• RQ3 Does combining multiple SVMs trained on different handcrafted descriptors with a CNN ensemble yield superior results compared to individual models?
• RQ4 To what extent can a general-purpose approach avoid overfitting when applied to unseen bioimage datasets?
• RQ5 What is the contribution of each component (handcrafted, learned, augmented data) to the overall performance gain?

Key findings

• The GenP ensemble achieves state-of-the-art performance on multiple bioimage classification benchmarks without any dataset-specific hyperparameter tuning.
• The proposed PCA/DCT-based data augmentation methods significantly improve model generalization and reduce overfitting.
• The combination of handcrafted descriptors and deep CNNs in an ensemble framework leads to higher accuracy than individual models.
• Each component of the ensemble—SVMs, CNNs, and data augmentation—contributes meaningfully to the final performance gain.
• The method demonstrates strong generalization across diverse biological image types, including microscopy and histopathology images.
• The absence of dataset-specific tuning reduces the risk of overfitting while maintaining high predictive accuracy.