Evaluating Performance of Multiclass Chest Radiograph Classification Using Soft Voting of CNNs and Transformers

Accurate and sensitive classification of multiple thoracic diseases using chest radiographs remains a significant challenge in computer-aided diagnosis.While Convolutional Neural Networks (CNNs) and Transformer-based models have shown strong performance individually is often limited.This study proposes a deep learning framework to improve the area under the curve (AUC) for multiclass detection of 14 chest diseases.We utilized the NIH ChestX-ray14 dataset, which includes 112,120 frontal-view chest radiographs from 30,805 unique patients, annotated with 14 disease labels.Predictions from five modelsincluding a CNN model (DenseNet121), two Transformer models (Vision Transformer and Swin Transformer), and two hybrid models (Hybrid ViT-DenseNet121 and Hybrid Swin-DenseNet121)-were integrated using soft voting to enhance classification robustness.Soft voting achieved an AUC of 0.8437, significantly higher than all individual or hybrid models (all P < 0.05), except DenseNet121 (AUC = 0.8433; P = 0.76).This framework significantly improves AUC for multiclass chest disease detection and offers a robust approach for more reliable and sensitive clinical screening applications.