Comparative Analysis of Deep Learning Frameworks for Multi-Label Chest X-Ray Classification
Nguyen Quang Minh
Department of Computer Science, Thai Nguyen University of Agriculture and Forestry
Tran Thi Hien
Department of Biomedical Engineering
Keywords: deep learning, convolutional neural network, chest x-ray, multi-label classification
Abstract
Chest x-rays are one of the most commonly performed radiological examinations for screening and diagnosis of various lung diseases. With advancements in deep learning, automated analysis of chest x-rays using convolutional neural networks (CNNs) has shown promise in improving radiological workflow. However, most existing studies have focused on single disease classification, while multi-label classification of comorbid thoracic diseases has been less explored. In this work, we perform a comparative analysis of popular deep learning frameworks - PyTorch, TensorFlow, and Keras with Tensorflow backend for multi-label classification of chest x-rays. We evaluate the frameworks on the NIH ChestX-ray14 dataset containing 112,120 x-ray images with 14 common thoracic disease labels. Pre-trained ResNet-50 is utilized as the base CNN architecture. The models are trained end-to-end with identical hyperparameters for a fair comparison. Evaluation metrics including AUC, precision, recall, F1-score, training speed, and model size are reported. Among the frameworks, TensorFlow achieves the best overall AUC of 0.9352, outperforming PyTorch (0.9201 AUC) and Keras (0.9114 AUC). However, PyTorch yields higher recall for minority labels like fibrosis and edema. Keras model has the fastest training speed and the smallest model size. The results demonstrate the strengths and weaknesses of each framework. Our findings serve as a reference to guide selection of deep learning frameworks for real-world deployment of multi-label chest x-ray classifiers. The Keras model offers a good speed-performance tradeoff while TensorFlow provides maximal discriminative ability.
Author Biography
Tran Thi Hien, Department of Biomedical Engineering
