Medical Image Analysis using Deep Convolutional Neural Networks: CNN Architectures and Transfer Learning

Over the past two decades, computer-aided detection or diagnosis (CAD) has been a fruitful area of research. Medical imaging technology can provide the radiologists and physicians with a more efficient diagnosis and treatment process through the medical image analysis. However, data analysis has slowly become a challenging task with the manual advancement of science and technology in the modern CAD systems. This problem can be successfully solved with the assistance of Deep Learning Methods, particularly convolutional neural networks (CNNs) for procedures such as breast cancer treatment, lung nodule detection and prostate cancer localization.Important progress has been made in image recognition, mainly due to the availability of large-scale annotated datasets and the revival of deeply convolutional neural networks (CNNs). CNNs may help to enable highly representative, data-driven, layered hierarchical image features to be learned from sufficient training data. CNNs can help enable highly representative, data-driven, layered hierarchical image features to be learned with sufficient training data. The three main strategies that successfully employ CNNs for medical image classification are currently training the CNN from scratch, using pre-trained CNN features off - theshelf and undertaking unattended pre-training with supervised fine tuning.An effective way to do this is by transfer learning, i.e. fine-tuning (supervised) pre-trained CNN models from the natural image dataset to medical image tasks (although the domain between two medical image datasets can also be transferred). Training is difficult if we are to train from scratch a profoundly convolutional neural network (CNN), because it requires a lot of labeled training data and a lot of experience to ensure proper convergence. To do this better, we can fine-tune a pre-trained CNN using, for example, a large set of natural images marked out.