Multi-Label Classification of Vindr-CXR Dataset Using Transfer-Learning Approach

Deep Learning (DL) based ComputerAided Diagnosis (CAD) systems have achieved great success in diagnosing common diseases in human body. However, the progression of health issues leads to occurrence of multiple nonexclusive diseases in medical images. This results into the problem of detection of Muli-Label (ML) diseases in diagnostic images. Existing ML disease classification techniques show promising results, but still have space for improvement. To solve this challenging task generally requires extraction of rich correlation information among disease labels. Therefore, this paper presents a collaborative framework, named as VMNet that improves the performance of ML classification by fusing features of best pre-trained models VGG19 and MobileNet. It explicitly explores image features from multi-label abnormalities existing in VinDr-CXR dataset using poweful technique of transfer learning. A systematic evaluation is performed by fine-tuning different models like MobileNet, ResNet50 and VGG19 on VinDr-CXR dataset. These models are investigated on the ROC-AUC performance metric. We observe that VGG19 and MobileNet networks gain the best AUC results of 86.4%, and 89.11% respectively compared to the ResNet50 network with an AUC score of 73.6 %. To maximize the classification efficiency and reduce execution complexity, collaborative framework VMNet fuses fine-tuned VGG19 and MobileNet models and achieves a better AUC value of 90.12% against state-of-the-art pipelines.