TOPSIS aided ensemble of CNN models for screening COVID-19 in chest X-ray images

Rishav Pramanik, Subhrajit Dey, Samir Malakar, Seyedali Mirjalili, Ram SarkarScientific 2022 · DOI

TL;DR

This paper introduces TOPCONet, a lightweight yet effective deep learning model for screening COVID-19 and pneumonia from chest X-rays. By combining three CNNs trained on different image variants (RGB, red channel, and edge-enhanced) and using the TOPSIS method for ensemble decision-making, TOPCONet achieves state-of-the-art performance with significantly fewer trainable parameters than existing models.

What Problem Does This Paper Address?

The COVID-19 pandemic highlighted the critical need for rapid and accurate diagnostic tools. While RT-PCR is the gold standard, it has limitations like low sensitivity and time delays. Radiological methods, particularly chest X-rays, offer a more accessible and faster alternative. However, interpreting these images can be challenging, leading to inter-observer variability and potential errors. Existing deep learning approaches, while promising, often require substantial computational resources and large numbers of trainable parameters, limiting their widespread adoption, especially in resource-constrained settings. This paper aims to bridge this gap by developing a computationally efficient and accurate Computer-Aided Diagnosis (CAD) system for COVID-19 and pneumonia detection from chest X-rays.

Methods

The core of the proposed method, TOPCONet, is an ensemble of three identical, lightweight Convolutional Neural Networks (CNNs). Each CNN acts as a "base classifier" and is trained on a different representation of the input chest X-ray image:

  • Classifier 1: Uses the standard RGB image.
  • Classifier 2: Uses only the red channel of the RGB image.
  • Classifier 3: Uses a 4-channel image created by combining the original RGB image with an edge-enhanced version derived from the Robert's edge detection method.

The rationale behind using these different input types is to capture complementary features. After these three base classifiers generate their probability scores for each class (COVID-19, Pneumonia, Normal), an ensemble strategy is employed. The authors utilize the Technique for Order Preference by Similarity to an Ideal Solution (TOPSIS), a multi-criteria decision-making (MCDM) method, to aggregate these scores. TOPSIS is chosen for its ability to rank alternatives based on multiple criteria and its non-trainable nature, which contributes to the overall efficiency of the model. The entire framework, TOPCONet, is designed to be "lightweight," meaning it requires fewer trainable parameters compared to many existing deep CNN models.

Datasets

The proposed TOPCONet model was evaluated on three publicly available datasets:

  1. Dataset-1: A combination of the IEEE COVID-19 dataset and the Kaggle Pneumonia Dataset.
  2. Dataset-2: The Kaggle Radiography Dataset.
  3. COVIDx Dataset: A large dataset specifically curated for COVID-19 detection.

The paper provides a detailed breakdown of the number of training and testing samples for each dataset in Table 1.

Key Insights and Ideas

  • Input Modality Matters: Using different representations of the same X-ray image (RGB, red channel, edge-enhanced) for separate CNNs significantly improves the ability to capture diverse and complementary features, leading to better overall performance.
  • TOPSIS for Ensemble Efficiency: The TOPSIS MCDM method proves to be an effective and computationally efficient way to combine the predictions of multiple CNNs. Its non-trainable nature avoids the need for additional data and computational resources often associated with trainable ensemble methods.
  • Lightweight is Right: The research demonstrates that achieving state-of-the-art performance doesn't necessarily require massive, parameter-heavy models. A carefully designed, lightweight CNN architecture, when combined with an effective ensemble strategy, can be highly competitive.
  • Edge Information is Valuable: Incorporating edge information (via Robert's edge detection) into the input data for one of the base classifiers proved beneficial, suggesting that subtle structural details in X-rays are important for accurate diagnosis.

Key Results

  • High Accuracy: TOPCONet achieved state-of-the-art results across the evaluated datasets. For instance, on Dataset-1, it achieved an accuracy of 98.78% (hold-out test set) and on Dataset-2, 98.61%. On the COVIDx dataset, it reported an accuracy of 94.25%.
  • Computational Efficiency: The model is significantly lighter than many existing CNN models. For example, TOPCONet has approximately 1 million trainable parameters, which is considerably less than models like VGG-19 (over 143 million) or ResNet101 (over 44 million).
  • Reduced Training Time: The lightweight nature of the model also translates to faster training times compared to many state-of-the-art transfer learning models.
  • Ensemble Advantage: The TOPSIS-aided ensemble consistently outperformed individual base classifiers and standard ensemble methods like majority voting, product rule, and sum rule. For example, on Dataset-1, it showed a 0.16% accuracy improvement over the second-best ensemble method.

Limitations and Future Work

  • Sensitivity to Image Variations: The authors acknowledge that the proposed CNN model might not perform optimally on rotated or arbitrarily cropped images and may require augmented training samples.
  • Ensemble Failure Scenario: A limitation of the TOPSIS ensemble is that it might fail to classify a sample correctly if two base classifiers incorrectly classify it with high confidence.
  • Resource Requirements: Despite being lightweight, deep learning models inherently demand resources like GPUs and RAM for training.
  • Future Work: The authors suggest focusing on handling imbalanced classes more efficiently, exploring domain knowledge integration to further boost performance, and investigating feature selection techniques to further reduce storage and time requirements. They also propose using the model to extract features for subsequent analysis.

Summary updated

Abstract

The novel coronavirus (COVID-19), has undoubtedly imprinted our lives with its deadly impact. Early testing with isolation of the individual is the best possible way to curb the spread of this deadly virus. Computer aided diagnosis (CAD) provides an alternative and cheap option for screening of the said virus. In this paper, we propose a convolution neural network (CNN)-based CAD method for COVID-19 and pneumonia detection from chest X-ray images. We consider three input types for three identical base classifiers. To capture maximum possible complementary features, we consider the original RGB image, Red channel image and the original image stacked with Robert's edge information. After that we develop an ensemble strategy based on the technique for order preference by similarity to an ideal solution (TOPSIS) to aggregate the outcomes of base classifiers. The overall framework, called TOPCONet, is very light in comparison with standard CNN models in terms of the number of trainable parameters required. TOPCONet achieves state-of-the-art results when evaluated on the three publicly available datasets: (1) IEEE COVID-19 dataset + Kaggle Pneumonia Dataset, (2) Kaggle Radiography dataset and (3) COVIDx.
BibTeX
@inproceedings{pramanik2022topsis,
  author = {Rishav Pramanik and Subhrajit Dey and Samir Malakar and Seyedali Mirjalili and Ram Sarkar},
  title = {{TOPSIS aided ensemble of CNN models for screening COVID-19 in chest X-ray images}},
  booktitle = {Scientific},
  year = {2022},
  doi = {10.1038/s41598-022-18463-7},
}
Figure