A Deep Learning Framework for Multi-Class Lung Disease Detection Using CXR Imaging

Technical Summary: A Deep Learning Framework for Multi-Class Lung Disease Detection Using CXR Imaging

Problem Statement
Early and accurate detection of lung diseases from Chest X-ray (CXR) images is critical for clinical decision-making. However, existing automated diagnostic systems face significant challenges, including overlapping visual patterns among different lung diseases (e.g., distinguishing between pneumonia, COVID-19, and lung opacity) and dataset bias. Models trained on multi-source datasets often learn "shortcut" features related to the data acquisition source (e.g., scanner type or hospital protocol) rather than genuine pathological features, leading to poor generalization in real-world clinical settings. Additionally, class imbalance in medical datasets often skews model performance toward majority classes, reducing sensitivity for underrepresented diseases.

Methodology
The authors propose a bias-aware deep learning framework designed for multi-class classification of six pulmonary conditions: COVID-19, Pneumonia, Tuberculosis, Lung Cancer, Lung Opacity, and Normal. The methodology is structured around three core pillars:

1. Bias-Aware Data Curation and Preprocessing:

   • Dataset Construction: A unified dataset of 13,825 CXR images was created by merging data from three public repositories (COVID-19 Radiography Database, IQ-OTH/NCCD Lung Cancer Dataset, and Kaggle repositories).
   • Source Harmonization: To eliminate directory-level cues, all images were stored in a single structure with a consolidated metadata file. Source labels were retained only for bias analysis, not for disease classification training.
   • Sanity Check: A "source prediction sanity check" was performed by training a ResNet-50 to predict the dataset origin. The model achieved only 51.57% accuracy with highly skewed recall, failing to reliably distinguish sources. This result supports the claim that the preprocessing successfully minimized dataset-specific acquisition artifacts, forcing the disease classification models to learn pathology-relevant features.
   • Class Balancing: To address class imbalance (where Lung Opacity and Pneumonia are dominant), the framework employed real-time data augmentation (rotation, flipping, zooming) and class-weighted loss functions during training.

2. Model Architecture and Training:

   • The study evaluated three ResNet architectures adapted for grayscale input: ResNet-38 (a customized, shallower variant derived from ResNet-34), ResNet-50, and ResNet-101.
   • All models utilized ImageNet-pretrained weights, modified to accept single-channel grayscale inputs via channel replication or layer adjustment.
   • Training employed 5-fold stratified cross-validation, dropout regularization (0.3), and the Adam optimizer with a weighted cross-entropy loss function.

3. Validation and Interpretability:

   • Statistical Validation: Performance was rigorously tested using bootstrap confidence intervals, Chi-square goodness-of-fit tests, and McNemar's tests for pairwise model comparison.
   • Interpretability: Grad-CAM (Gradient-weighted Class Activation Mapping) was used to visualize model attention. A qualified medical expert (Dr. Govind Saran) qualitatively validated that the heatmaps focused on anatomically plausible lung regions rather than background artifacts.
   • External Validation: The best-performing model was tested on an independent, unseen dataset from Roboflow (800 images) without retraining to assess generalization.

Key Contributions

1. Bias-Aware Framework: The paper introduces a specific preprocessing pipeline that includes a "source prediction sanity check" to explicitly verify that models are not learning dataset-origin shortcuts.
2. Architectural Insight (Depth vs. Generalization): Contrary to the trend of increasing network depth, the study demonstrates that a shallower architecture (ResNet-38) outperformed deeper counterparts (ResNet-50 and ResNet-101) on this specific multi-class, class-imbalanced medical dataset.
3. Comprehensive Evaluation: The study provides a robust statistical validation suite (including McNemar tests and bootstrap analysis) and external validation on an unseen dataset, moving beyond simple accuracy metrics to assess reliability and generalization.

Experimental Results

• Performance: ResNet-38 achieved the highest mean classification accuracy of 93.34% (±0.0073) under 5-fold cross-validation. This outperformed ResNet-50 (86.97%) and ResNet-101 (89.40%).
• Stability: ResNet-38 demonstrated lower variance across folds and better calibration (more consistent posterior probabilities) compared to the deeper models, which showed signs of overfitting (training accuracy approaching 1.0 while validation lagged).
• Class Performance: ResNet-38 showed superior F1-scores and recall across most classes, particularly for minority classes like Tuberculosis and COVID-19, attributed to the effective handling of class imbalance.
• External Validation: On the unseen Roboflow dataset, the ResNet-38 model achieved 85% accuracy, with errors primarily occurring between clinically overlapping conditions (e.g., Pneumonia vs. COVID-19), confirming the model learns pathological features rather than dataset artifacts.
• Efficiency: ResNet-38 required significantly fewer parameters (21M) and lower computational resources (FLOPs ~3.1G) compared to ResNet-101 (44.5M parameters, ~7.8G FLOPs), offering a better efficiency-accuracy trade-off.

Significance and Claims
The paper claims that for multi-class lung disease classification using CXR images, increasing network depth does not necessarily improve performance, especially when dealing with moderate-sized, class-imbalanced datasets. The study argues that deeper networks (like ResNet-101) are prone to overfitting and learning noise, whereas a carefully optimized, shallower architecture (ResNet-38) combined with rigorous bias-mitigation strategies yields superior generalization and stability.

The authors emphasize that their framework provides a "realistic method for AI-based radiographic screening" by prioritizing bias-aware design and statistical validation over mere architectural complexity. They explicitly state that while the framework reduces dataset-origin dependency, it does not eliminate all domain shifts, and further external validation on diverse hospital datasets is required for clinical deployment. The work concludes that the ResNet-38 model represents an optimal balance between accuracy, computational efficiency, and reliability for real-world CXR analysis.