Transfer learning methods for classifying medical images and identifying illnesses with minimal annotated data

Abstract

Medical image classification is often constrained by the unavailability of large, annotated data-sets owing to the unreasonable cost of expert labeling and class discrepancy in rare conditions. One of the interesting techniques to overcomethesedrawbacksistheprocedureoftransfer learning (TL) which utilizes the knowledge gained from deep learning (DL) models that were formerly trained. DL has achieved significant success in feature representation from complex data, but it generally involves a considerable amount of annotated information for effective training. Using small and imbalanced data-sets, this paper investigates the performance of TL for medical image classification tasks using VGG16, ResNet50, and EfficientNet. A scratch-built baseline Convolutional Neural Network (CNN) is adopted as a reference to compare TL models. COVID-19 chest X-ray databases are utilized in experiments. Accuracy, F1-score, and training efficiency are applied to measure performance. While TL has illustrated impressive performance, often similar to or exceeding diagnosis, its configuration in numerous studies has been defined as arbitrary. Previous studies have revealed that TL is better than scratch-built models on tiny medical datasets, resulting in a COVID-19 detection precision of >96% using ResNet50. Also pointed out the usefulness of ensemble TL approaches. This paper shows that TL is a practicable and effective means to expand medical image solutions in cases with limited data. That makes it probable to quickly deploy artificial intelligence (AI) tools into real-world healthcare applications