Effectiveness of Transfer Learning on Medical Image Classification Using Chest-Xray 14 Dataset
Transfer learning is significantly gaining rapid adoptions as an important tool for diagnosis and interpretation of medical images by decreasing the time spent in predictions, improving the accuracy in identifying abnormalities and, therefore, enhancing the clinical outcomes of patients. We test the effectiveness of transfer learning (TL) techniques, namely, transferring knowledge from deep learning models pre-trained with general-purpose images to medical image classification using the ChestX-ray 14 dataset, comprising of 112,120 frontal-view chest X-ray images from 30,805 unique patients. We use the DenseNet-121 architecture, pre-trained on ImageNet, as our baseline model and perform a binary classification on our dataset. The results show that fine-tuning with data augmentation gives a more robust model performance and we propose that identifying the optimal cut-off layer during fine-tuning provides a novel approach for higher-order representation of medical features. For future research, we will combine fine-tuning approaches with hyperparameter optimization, adding non-image patient data, finding optimal data augmentation and model architecture, and generating high resolution medical images using generative adversarial networks to improve model performance.
Index Terms - Chest X-ray, Deep Learning, Medical Image Classification, Transfer Learning.