MEDNC: Multi-ensemble deep neural network for COVID-19 diagnosis

Lin Yang, Shuihua Wang, Yudong ZhangarXiv 2023 · DOI · arXiv

Introduction

The global pandemic necessitated rapid and accurate diagnostic tools. While Reverse Transcription Polymerase Chain Reaction (RT-PCR) remains the clinical gold standard, it is often hampered by limited sensitivity and potential false negatives. Chest Computed Tomography (CT) scans offer a high-sensitivity alternative, but manual interpretation is time-consuming and requires specialized expertise.

Existing AI-based diagnostic tools face significant hurdles:

  • Data Scarcity: Training robust models from scratch requires vast amounts of labeled medical data.
  • Model Variance: Individual architectures often yield inconsistent results across different datasets.
  • Data Imbalance: Many studies suffer from biased datasets, leading to inflated performance metrics.
Example of a COVID-19 CT scan showing characteristic lung abnormalities.

The MEDNC Framework

To address these challenges, researchers developed the Multi-ensemble Deep Neural Network (MEDNC). This framework leverages transfer learning by utilizing 16 popular CNN architectures (e.g., ResNet, DenseNet, VGG, MobileNet) pre-trained on the ImageNet dataset. By "freezing" feature extraction layers and replacing classifiers with custom networks, the model benefits from a broad knowledge base of visual features.

Hierarchical Ensemble Strategy

The core innovation lies in the multi-layer ensemble approach. Models are combined at various depths to capture diverse information:

  • FFC-MEDNC: Ensembles at the Feature and Fully Connected layers.
  • FO-MEDNC: Ensembles at the Feature and Output layers.
  • FCO-MEDNC: Ensembles at the Fully Connected and Output layers.
  • FFCO-MEDNC: A comprehensive "Level 3" ensemble combining all three—Feature, Fully Connected, and Output layers.
The MEDNC workflow, illustrating the path from dataset splitting to multi-ensemble classification.

Experimental Results

The framework was validated using Monte Carlo Cross-Validation (MCCV) across four datasets: two COVID-19 CT scan benchmarks, a brain tumor X-ray dataset, and a blood cell microscopic image dataset.

Performance and Generalizability

The FFCO-MEDNC variant consistently outperformed individual models and simpler ensembles:

  • COVID-19 Diagnosis: Achieved 98.79% accuracy on Dataset A and 99.82% on Dataset B.
  • Cross-Domain Utility: Demonstrated remarkable generalizability with 99.39% accuracy for brain tumor detection and 99.28% for blood cell classification.
  • Efficiency: Despite its complexity (392.5 MB), the model maintains a training time of approximately 31 seconds per epoch on consumer-grade hardware (RTX 2070S), making it viable for clinical settings.

Conclusion and Future Directions

MEDNC represents a significant step forward in automated medical diagnosis. By ensembling diverse architectures at multiple levels, the framework achieves state-of-the-art accuracy while remaining versatile enough for various imaging tasks. Future research will focus on automating the selection of 2D slices from 3D CT volumes and integrating advanced image enhancement techniques to further refine diagnostic precision.

Summary updated

BibTeX
@misc{yang2023mednc,
  author = {Lin Yang and Shuihua Wang and Yudong Zhang},
  title = {{MEDNC: Multi-ensemble deep neural network for COVID-19 diagnosis}},
  year = {2023},
  eprint = {2304.13135},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2304.13135},
}
Figure