[Paper Review] Efficient and Accurate Pneumonia Detection Using a Novel Multi-Scale Transformer Approach

Ahmed Saber, Amirreza Fateh+4 more|arXiv (Cornell University)|Aug 8, 2024

Advanced Chemical Sensor Technologies4 citations

TL;DR

This paper proposes a novel multi-scale Transformer-based method for efficient and accurate pneumonia detection from chest X-rays, combining TransUNet for lung segmentation with a lightweight transformer module applied to feature maps from ResNet backbones. The approach achieves 92.79% accuracy on the Kermany dataset and 95.11% on the Cohen dataset, significantly improving performance with fewer parameters than standard transformers.

ABSTRACT

Pneumonia, a prevalent respiratory infection, remains a leading cause of morbidity and mortality worldwide, particularly among vulnerable populations. Chest X-rays serve as a primary tool for pneumonia detection; however, variations in imaging conditions and subtle visual indicators complicate consistent interpretation. Automated tools can enhance traditional methods by improving diagnostic reliability and supporting clinical decision-making. In this study, we propose a novel multi-scale transformer approach for pneumonia detection that integrates lung segmentation and classification into a unified framework. Our method introduces a lightweight transformer-enhanced TransUNet for precise lung segmentation, achieving a Dice score of 95.68% on the "Chest X-ray Masks and Labels" dataset with fewer parameters than traditional transformers. For classification, we employ pre-trained ResNet models (ResNet-50 and ResNet-101) to extract multi-scale feature maps, which are then processed through a modified transformer module to enhance pneumonia detection. This integration of multi-scale feature extraction and lightweight transformer modules ensures robust performance, making our method suitable for resource-constrained clinical environments. Our approach achieves 93.75% accuracy on the "Kermany" dataset and 96.04% accuracy on the "Cohen" dataset, outperforming existing methods while maintaining computational efficiency. This work demonstrates the potential of multi-scale transformer architectures to improve pneumonia diagnosis, offering a scalable and accurate solution to global healthcare challenges. https://github.com/amirrezafateh/Multi-Scale-Transformer-Pneumonia

Motivation & Objective

• To address the challenge of inconsistent pneumonia diagnosis in underdeveloped regions due to variability in radiologist interpretation.
• To improve diagnostic accuracy and efficiency in pneumonia detection using deep learning and attention mechanisms.
• To reduce model complexity and parameter count in Transformer-based medical image classification while maintaining high performance.
• To enhance classification by integrating lung segmentation to focus on relevant anatomical regions and reduce noise.
• To develop a deployable, efficient solution suitable for resource-constrained clinical environments.

Proposed method

• Lung segmentation is performed using a TransUNet model enhanced with a novel, parameter-efficient transformer module trained on the Chest Xray Masks and Labels dataset.
• The segmented lung regions are extracted and used as input to fine-tune pre-trained ResNet-50 and ResNet-101 models for multi-scale feature extraction.
• A modified, lightweight transformer module processes the multi-scale features to model long-range dependencies and emphasize critical patterns in the lung regions.
• The model employs transfer learning by initializing with ImageNet-pretrained weights and fine-tuning on the Kermany and Cohen pneumonia datasets.
• The integration of segmentation and attention mechanisms reduces background noise and improves feature relevance for classification.
• The proposed method is evaluated via ablation studies comparing performance on raw images versus segmented image inputs.

Figure 1 : the overview of proposed method

Experimental results

Research questions

• RQ1 Can a lightweight, multi-scale Transformer module improve pneumonia classification accuracy while reducing parameter count compared to standard Transformers?
• RQ2 How does lung segmentation using TransUNet affect the performance of subsequent pneumonia classification models?
• RQ3 To what extent does the integration of attention mechanisms in a segmented, multi-scale feature framework enhance diagnostic accuracy on chest X-ray datasets?
• RQ4 Does the proposed method achieve superior performance with fewer parameters than existing state-of-the-art Transformer-based models for pneumonia detection?
• RQ5 Can this approach maintain high accuracy in resource-constrained environments due to its low parameter count and efficient design?

Key findings

• The proposed method achieved 92.79% accuracy on the Kermany dataset and 95.11% accuracy on the Cohen dataset, demonstrating strong generalization and robustness.
• Incorporating lung segmentation masks improved classification accuracy by 2 percentage points for ResNet-50 and 2.73 percentage points for ResNet-101.
• The model achieved higher precision, recall, F1-score, and AUC when using segmented images compared to raw images, indicating reduced misclassification.
• The ablation study confirmed that segmentation significantly reduces noise and enhances feature relevance, leading to consistent performance gains across both ResNet backbones.
• The proposed transformer module maintains high performance with fewer parameters than standard Transformers, making it suitable for deployment in low-resource settings.
• Confusion matrices showed a consistent reduction in false positives and false negatives when using predicted lung masks, particularly in the pneumonia and COVID-19 categories.

Figure 2 : the overview of transformer