Related work

replicated the base CheXNet (Rajpurkar et al., 2017) on NIH :)Strick, Garcia, & Huang, 2025(
ChestX-ray14 data and introduced an improved model, DannyNet. Incorporating Focal Loss,
AdamW optimizer, best-practice augmentations, and per-class F1 thresholds, DannyNet
performed better than their CheXNet replica (AUC 0.79, F1 0.08) and a Vision Transformer
baseline with better performance (AUC 0.85, F1 0.39). The effort emphasizes reproducibility
with public code and a web application of Grad-CAM visualizations for interpretability.
However, there remain performance gaps: mean F1 values are not yet impressive, Vision
Transformers did not do well on the dataset, and comparisons to radiologists are hindered by
the lack of publicly available expert-labeled test sets. The paper illustrates how small training
optimizations can significantly improve model reliability, as well as emphasizes the persistent
problems of class balance, precision-recall trade-offs, and clinical applicability in AI for
.medical imaging

Critically evaluate the impact of common design :)Vogelbaum, Engstrom, & Madry, 2023(
choices on chest X-ray classification, e.g., architectures (ResNet, DenseNet, VGG), loss
functions (cross-entropy, focal, hierarchical, AUROC-based), pooling strategies (standard vs.
PCAM), and data augmentations. With three baselines—CheXpert, MIMIC-CXR, and ChestX-
ray14—they find that special-purpose methods rarely outperform general baselines. A vanilla
ResNet-50 with binary cross-entropy and no data augmentation is competitive to, or better
than, most hand-crafted models, with improvements from other alternatives usually less than
2% AUROC. Most interestingly, standard methods such as DenseNet-121 and hierarchical
.losses yielded small or erratic improvements

The authors hypothesize two possible reasons: (1) effective domain-specific strategies have
not yet been formulated, or (2) current benchmarks—with noisy labels and low clinical fidelity
—fail to exhibit improvements. The study highlights the necessity of robust baselines,
estimation of uncertainty, and multi-dataset evaluation before proclaiming breakthroughs in
.medical image AI

In this work, ChestXray8 is introduced, a vast dataset of over 100,000 :)Wang, et al., 2017(
chest X-rays of nearly 33,000 patients. The disease labels for each image (e.g., Atelectasis,
Pneumonia) were automatically extracted from radiology reports using Natural Language
.Processing (NLP)

The authors demonstrate how these images can be leveraged to train a deep learning model
for weakly-supervised multi-label disease localization and classification. This means the
model can learn to detect and crudely locate diseases in the X-rays using image-level labels
only (i.e., whether a disease is present or not), without location information. The outcomes are
encouraging but also reveal that high-accuracy, autodriven diagnosis remains a grand
.challenge. The database is a valuable public benchmark for advancing AI in medical imaging
This comprehensive review discusses the application of deep :)Siddiqi & Javaid, 2024(
learning to identify pneumonia from chest X-ray images, on the one hand, on the extremely
encouraging advancements, as well as the most important difficulties in the field. Research
findings from 2020-2023 indicate that CNNs and transfer learning techniques have achieved
accurate diagnosis with prospective solutions to address the global radiologist shortage.
However, the study highlights some of the major clinical adoption barriers including dataset
bias, model interpretability issues, vulnerability to adversarial attacks, and class imbalance in
training. The authors deduce that vision transformers (ViTs) are the most promising future
direction due to their potential for rich context understanding of images, though general
robustness and transparency issues need to be addressed before the systems can be made
.reliable enough to apply them in medical practice

Materials and Methods

Datasets or samples: The study employed the publicly available NIH ChestX-ray14 dataset,
developed by the U.S. National Institutes of Health (NIH) Clinical Center. The dataset
contains 112,120 frontal chest X-ray images from 30,805 patients across 14 thoracic disease
classes with corresponding image-level labels derived from radiology reports by natural
.language processing (NLP)

For this project, a group of eight significant pathologies was selected to focus on the most
clinically significant and well-studied conditions: Atelectasis, Cardiomegaly, Effusion,
.Infiltration, Mass, Nodule, Pneumonia, and Pneumothorax

Tools and libraries: The software stack included

Python 3.10

TensorFlow 2.15

Keras API

Scikit-learn 1.5

.NumPy, Pandas, Matplotlib, and Seaborn for data handling and visualization

Hardware setup: All experiments were performed in Kaggle's GPU environment on an NVIDIA
.Tesla T4 GPU (16 GB VRAM)

The proposed model

We proposed that a comprehensive deep learning model utilizes a robust approach for the
multi-label classification of thoracic pathologies on chest X-rays. The architecture begins with
loading and preprocessing the NIH ChestX-ray14 dataset. Preprocessing pipeline consists of
EfficientNet-specific normalization and pipelined data loading optimized through TensorFlow's
.data API to effectively process high-volume medical image data
The model architecture integrates several cutting-edge features: an EfficientNet-B0 ImageNet
pre-trained backbone for robust feature extraction, a convolutional attention block learning to
pay attention to diagnostically informative regions, and an optimal global average and max
pooling strategy for global feature representation. The training employs a highly sophisticated
two-stage approach with focal loss for tackling extreme class imbalance, followed by
.Stochastic Weight Averaging (SWA) for optimizing generalization and convergence stability

The performance framework provides comprehensive assessment on various metrics

including ROC-AUC, PR-AUC, F1 scores, and optimal thresholds specific to pathology. The
model also offers extensive visualization like ROC curves, precision-recall curves, and
confusion matrices for each pathology, which can help enable comprehensive clinical
.validation and model interpretability

The proposed model block diagram

Results and discussion

The model was evaluated on the validation subset containing 10,824 chest X-ray images.
.Performance metrics were computed per pathology, as shown in table 1

Optimal Threshold Recall Precision F1-score PR-AUC ROC-AUC Pathology

0.1745 0.6769 0.0604 0.1108 0.1310 0.7939 Cardiomegaly

0.2248 0.7701 0.1862 0.2999 0.2427 0.7586 Atelectasis

0.2036 0.8219 0.2527 0.3866 0.4214 0.8462 Effusion

0.2448 0.6572 0.2245 0.3347 0.2764 0.6655 Infiltration

0.1840 0.5631 0.1065 0.1791 0.1330 0.7201 Mass

0.1918 0.4473 0.0954 0.1572 0.1225 0.6523 Nodule

0.1114 0.4833 0.0209 0.0402 0.0234 0.6590 Pneumonia

0.1553 0.7644 0.0694 0.1272 0.1435 0.7779 Pneumothorax

Performance Comparison with Popular Public Implementations

Training Cardiomegaly ROC- Effusion ROC- Avg PR- Avg ROC-

Source Key Features Model / Approach
Time AUC AUC AUC AUC

EfficientNet + Attention + Focal

This Work hours 2.1~ 0.794 0.846 0.187 0.759 Proposed Approach
Loss + SWA

Rajpurkar et al.,[ CheXNet (DenseNet-

DenseNet-121 + Class Balancing hours 3.5~ 0.768 0.824 0.162 0.741
]2017 121)

]NIH Official[ Standard ResNet-50 hours 2.8~ 0.751 0.812 0.148 0.728 Baseline ResNet-50

EfficientNet-B4 + Basic EfficientNet-B4

]Kaggle Public[ hours 4.2~ 0.782 0.838 0.175 0.752
Augmentation Kaggle

]Colab Public[ Custom CNN + Weighted Loss hours 2.5~ 0.760 0.819 0.155 0.735 Multi-label CNN

Transfer Learning
]Kaggle Notebook[ VGG16 + Fine-tuning hours 3.0~ 0.743 0.805 0.142 0.719
VGG16

Kaggle[
Multiple Model Ensemble hours 6.0~ 0.775 0.832 0.169 0.748 Ensemble Approach
]Competition
References

Ge, Z., Mahapatra, D., Sedai, S., Chakravorty, R., & Garnavi, R. (2018). Chest X-rays
Classification a Multi-Label and Fine-Grained Problem. arXiv, 1 -9.

Hossain, E. (2024). Machine Learning Crash Course for Engineers. Cham, Switzerland:
Springer.

Kapavarapu, S. K. (2025). NIH-Chest-X-rays-Multi-Label-Image-Classification.

[Link], 1 -9.

Li, X., Xu, X., Liu, Y., & Zhao, X. (2025). CheX-DS: Improving Chest X-ray Image
Classification with Ensemble Learning Based on DenseNet and Swin Transformer.
arXiv, 1- 7.

Nasser, A. A., & Akhloufi, M. A. (2023). A Review of Recent Advances in Deep Learning
Models for Chest Disease Detection Using Radiography. MDPI- Diagnostics , 1 -36.

Pooch, E. H., Ballester, P. L., & Barros, R. C. (2020). Can we trust deep learning based
diagnosis? The impact of domain shift in chest radiograph classification. arXiv, 1 -10.

Rajpurkar, P., Irvin, J., Zhu, K., Yang, B., & Mehta, H. (2017). CheXNet: Radiologist-Level
Pneumonia Detection on Chest X-Rays with Deep Learning. arXiv, 1 -7.

Siddiqi, R., & Javaid, S. (2024). Review Deep Learning for Pneumonia Detection in Chest X-
ray Images: A Comprehensive Survey. Journal of Imaging, 1- 35.

Strick, D., Garcia, C., & Huang, A. (2025). Reproducing and Improving CheXNet: Deep
Learning for Chest X-ray Disease Classification. arXiv, 1-12.

Vogelbaum, E., Engstrom, L., & Madry, A. (2023). What Works in Chest X-Ray Classification?
A Case Study of Design Choices. Workshop on Interpretable ML in Healthcare at
International Conference on Machine Learning (ICML) (pp. 1-22). Honolulu:
[Link]

Wang, X., Peng, Y., Lu, L., Lu, Z., Bagheri, M., & Summers, R. M. (2017). ChestX-ray8:
Hospital-scale Chest X-ray Database and Benchmarks on Weakly-Supervised
Classification and Localization of Common Thorax Diseases. arXiv, 1 - 19.