🤖 AI

Comparative Analysis of Deep Learning Architectures for Multi-Disease Classification of Single-Label Chest X-rays

This study systematically evaluates seven deep learning architectures on a balanced dataset of 18,080 chest X-rays, demonstrating that ConvNeXt-Tiny achieves the highest accuracy (92.31%) while MobileNetV2 offers the best parameter efficiency, with all models exceeding 90% accuracy and showing near-perfect performance for Tuberculosis and COVID-19 classification.

Original authors: Ali M. Bahram, Saman Muhammad Omer, Hardi M. Mohammed

Published 2026-03-17

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Ali M. Bahram, Saman Muhammad Omer, Hardi M. Mohammed

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a world where a doctor's most powerful tool for checking your lungs and heart is a simple X-ray picture. It's cheap, fast, and everywhere. But here's the catch: reading these pictures is hard. It requires a highly trained expert (a radiologist), and even experts can get tired, make mistakes, or disagree with each other. In many parts of the world, there simply aren't enough experts to go around.

This paper is like a grand tasting competition for seven different "AI chefs" (computer programs) to see which one is best at reading these X-ray pictures to diagnose five different problems: an enlarged heart, COVID-19, normal lungs, pneumonia, and tuberculosis.

Here is the breakdown of what they did and what they found, using some everyday analogies.

1. The Setup: Building a Fair Playground

Before the race could start, the researchers had to fix a major problem with the "ingredients" (the data).

The Problem: They had a huge pile of "Normal" X-rays (like having 60,000 apples) but very few pictures of sick people (like only 1,000 oranges). If you train a chef with mostly apples, they will just guess "apple" every time and get a high score, but they'll fail when an orange shows up.
The Fix: They used a technique called "Balancing the Scale." They took a huge chunk of the "Normal" pictures out (like putting some apples in a box) and used a digital "photocopier" (data augmentation) to create more pictures of the sick patients. Now, they had an equal number of pictures for every disease. This ensured the AI learned to spot the disease, not just the lack of disease.

2. The Contestants: Seven Different AI Architects

They invited seven famous AI models to compete. Think of them as different types of detectives, each with a unique style of looking at clues:

The Old Reliables (ResNet50, DenseNet121/201): These are like veteran detectives. They've been around for years, are very thorough, and have a proven track record.
The Speedsters (MobileNetV2): This is the lightweight detective. It's small, fast, and designed to run on a smartphone. It's like a scout who can run up a mountain quickly without a heavy backpack.
The Modernizers (ConvNeXt-Tiny, EfficientNet): These are the new kids on the block. They take the best ideas from old detectives and mix them with new, high-tech gadgets. They are trying to be the perfect blend of speed and smarts.
The Visionaries (ViT-B/16): This is the "Transformer" detective. Instead of looking at clues one by one, it looks at the whole picture at once to understand the big context. It's like a detective who reads the entire crime scene report before looking at a single fingerprint. However, it's also the heaviest and most expensive to run.

3. The Race: How They Performed

All seven detectives were given the exact same training, the same test questions, and the same rules. Here's how they fared:

The Gold Medalist (ConvNeXt-Tiny): This model won the race for accuracy. It got the right answer about 92% of the time. It proved that you don't need to be the biggest or oldest detective to be the smartest; you just need the right modern tools.
The Runner-Ups (ResNet50 & DenseNet): The veteran detectives came in a very close second and third. They were incredibly reliable, proving that the "old school" methods are still very strong.
The Efficiency Champion (MobileNetV2): This is the most important finding for the real world. While it wasn't the absolute fastest at getting the perfect score, it was the most efficient.
- The Analogy: Imagine the Gold Medalist is a Ferrari. It's fast and precise, but it guzzles gas and costs a fortune to maintain. MobileNetV2 is a hybrid Toyota Prius. It gets you to the destination almost as fast (90%+ accuracy), but it uses 87% less fuel (computer power) and fits in a tiny garage (runs on a phone).
- Why this matters: In a remote village with no internet or expensive computers, you can't run the Ferrari. You need the Prius. This model showed you can get excellent results without needing a supercomputer.

4. The Weak Spots: Where the AI Got Confused

Even the best detectives make mistakes. The AI was almost perfect at spotting Tuberculosis and COVID-19 (like spotting a bright red fire hydrant in a white room).

However, it struggled a bit with:

Normal vs. Enlarged Heart: Sometimes a heart looks a little big, but is it actually sick? It's a gray area, even for human doctors.
Normal vs. Pneumonia: Early pneumonia can look like a faint smudge that might just be a shadow.
The Lesson: The AI isn't failing; it's facing the same ambiguity that human doctors face. It's not a bug; it's a feature of the complex human body.

5. The "X-Ray Vision" Check (Grad-CAM)

To make sure the AI wasn't just guessing randomly, the researchers used a tool called Grad-CAM. Think of this as putting a glowing highlighter on the X-ray to show where the AI was looking.

When it said "Pneumonia," the highlighter glowed on the infected part of the lung.
When it said "Enlarged Heart," it glowed on the heart border.
This proved the AI was actually "seeing" the disease, not just guessing based on the background.

The Big Takeaway

The main message of this paper is hope and practicality.

We used to think that to get a super-smart medical AI, you needed a massive, expensive supercomputer and a huge, perfect dataset. This study shows that you don't.

You can build a system that is nearly as accurate as the best super-computer, but it's small enough to run on a laptop or a phone. This means that in the future, a nurse in a remote clinic could take an X-ray, run it through a lightweight app (like MobileNetV2), and get a highly accurate diagnosis in seconds, saving lives where doctors are scarce.

In short: We found the "Goldilocks" AI—not too big, not too small, but just right for saving lives everywhere.

1. Problem Statement

Chest X-ray (CXR) imaging is the primary diagnostic tool for pulmonary and cardiac disorders due to its low cost and accessibility. However, manual interpretation suffers from inter-observer variability, radiologist shortages, and fatigue, leading to delayed or missed diagnoses. While Deep Learning (DL) has shown promise, existing literature often lacks rigorous, controlled comparisons of contemporary architectures (CNNs vs. Transformers) on balanced, multi-class datasets. Most studies utilize imbalanced data (where "Normal" cases vastly outnumber diseases) or test single architectures under varying conditions, making it difficult to determine the true trade-offs between diagnostic accuracy and computational efficiency for resource-constrained healthcare settings.

2. Methodology

Dataset Construction and Preprocessing

Data Sources: The authors constructed a unified dataset by integrating three public repositories: NIH Clinical Center (Normal, Pneumonia, Cardiomegaly), COVID-19 Radiography Database, and Tuberculosis Chest X-ray Database.
Classes: Five single-label classes: Cardiomegaly, COVID-19, Normal, Pneumonia, and Tuberculosis.
Balancing Strategy: To address severe class imbalance (e.g., 60,361 Normal vs. 1,390 Pneumonia images), the authors employed:
- Random Undersampling: Reduced the majority "Normal" class to match the largest minority class (3,616 images).
- Controlled Oversampling: Applied data augmentation (horizontal flipping, ±15° rotation, brightness/contrast adjustments) to minority classes to reach 3,616 images each.
Patient-Level Splitting: To prevent data leakage, the dataset (18,080 images from 6,978 unique patients) was split at the patient level (not image level) into:
- Training: 70% (14,563 images)
- Validation: 10% (1,455 images)
- Testing: 20% (3,517 images)
- Note: No patient appeared in more than one split.

Experimental Framework

Architectures Evaluated: Seven state-of-the-art models were fine-tuned using transfer learning (ImageNet pre-trained weights) under identical conditions:
1. DenseNet121 & DenseNet201: Dense connectivity for feature reuse.
2. ResNet50: Residual learning to mitigate vanishing gradients.
3. MobileNetV2: Lightweight, inverted residual blocks for efficiency.
4. EfficientNetV2-M: Compound scaling for optimized depth/width/resolution.
5. ConvNeXt-Tiny: Modern CNN incorporating Transformer design principles (large kernels, LayerNorm).
6. ViT-B/16: Vision Transformer using self-attention mechanisms.
Training Configuration:
- Optimizer: Adam ( $\beta_1=0.9, \beta_2=0.999$ ).
- Learning Rate: Fixed at $1e^{-4}$ .
- Loss Function: CrossEntropyLoss.
- Augmentation: Standardized pipeline (flips, rotations, brightness).
- Inference: TenCrop evaluation with probability averaging.
Evaluation Metrics: Accuracy, Precision, Recall, F1-Score, AUROC (Area Under the Receiver Operating Characteristic Curve), and computational efficiency (parameters, training/inference time).

3. Key Contributions

Balanced Multi-Class Dataset: Creation of a rigorously balanced 5-class CXR dataset (3,616 images per class) to eliminate bias toward majority classes, ensuring fair evaluation of minority diseases.
Systematic Comparative Study: A head-to-head comparison of seven diverse architectures (classical CNNs, modern CNNs, and Transformers) under strictly standardized training protocols, addressing the lack of direct comparisons in current literature.
Efficiency-Accuracy Trade-off Analysis: Introduction of an "AUROC per Parameter" metric to quantify architectural efficiency, identifying models that offer high performance with minimal computational cost.
Clinical Interpretability: Application of Grad-CAM to visualize model attention, confirming that the models focus on clinically relevant anatomical regions (e.g., lung parenchyma for pneumonia, cardiac silhouette for cardiomegaly).

4. Key Results

Performance Overview

All seven models achieved test accuracies exceeding 90%, demonstrating the viability of DL for multi-disease classification.

Top Performer (Accuracy): ConvNeXt-Tiny achieved the highest test accuracy (92.31%) and the highest validation AUROC (98.64%). It effectively bridges the gap between CNN efficiency and Transformer-like feature extraction.
Runner-ups: ResNet50 (92.00%) and ViT-B/16 (91.87%) followed closely.
Efficiency Champion: MobileNetV2 achieved 90.42% test accuracy with only 3.5 million parameters (87.5% fewer than the best model) and the fastest inference time (1m 20s).

Computational Efficiency

MobileNetV2 demonstrated the highest efficiency ratio (AUROC per parameter), making it ideal for mobile or edge devices.
ViT-B/16, despite having the most parameters (85.8M), did not outperform smaller CNNs, suggesting that Transformers require larger datasets or domain-specific pre-training to leverage their global attention mechanisms effectively in medical imaging.

Disease-Specific Insights

High Performance: Tuberculosis and COVID-19 were classified with near-perfect accuracy (AUROC $\ge$ 99.97%) due to distinct radiographic features (e.g., cavitary lesions, ground-glass opacities).
Challenging Classes: Normal, Cardiomegaly, and Pneumonia showed lower accuracy (90–92%) and higher confusion rates. This is attributed to overlapping features (e.g., distinguishing early pneumonia from normal vascular markings or borderline cardiomegaly).

Convergence

CNN-based models (DenseNet, MobileNet, ConvNeXt) converged efficiently (78–86% of training epochs).
ViT-B/16 required the full 50 epochs to converge, indicating slower learning dynamics on this dataset size.

5. Significance and Conclusion

This study provides critical evidence that high-accuracy multi-disease classification is achievable without massive computational resources.

Resource-Constrained Settings: The findings support the deployment of lightweight models like MobileNetV2 in rural clinics or mobile health applications where hardware is limited.
High-Resource Settings: For maximum diagnostic precision, ConvNeXt-Tiny is recommended as the optimal balance of modern design and performance.
Clinical Implications: The study highlights that while AI excels at distinct pathologies (TB, COVID-19), challenges remain in differentiating subtle or overlapping conditions (Normal vs. Cardiomegaly). Future work should focus on multi-label learning, integrating clinical metadata, and multi-center validation to further improve robustness.

In summary, the paper successfully demonstrates that architectural choice significantly impacts the trade-off between diagnostic fidelity and computational cost, offering a roadmap for deploying AI diagnostics in diverse global healthcare environments.