More efficient deep learning for medical image analysis

深度学习形式的人工智能，例如使用卷积神经网络，对医学图像分析产生了巨大影响。它在会议和期刊出版物中占据主导地位，并在许多基准测试和应用中展示了最先进的性能，在某些情况下优于人类观察者。但是，尽管如此，在常规临床实践中采用这些方法的速度非常缓慢。其中一个原因是深度学习模型的训练效率低下且成本高昂，通常需要数万或数十万张专业标记的训练图像，并且需要在高端GPU硬件上进行许多天的训练。对于医疗应用来说，需要如此多的专业标记数据是一个关键挑战。毕竟，放射科医生（经过专门训练来解释医学图像的医生）能够使用更小的训练图像集来学习新任务。该项目将研究提高深度学习模型训练效率的方法，减少所需训练集的大小和/或细节水平，同时保持诊断准确性。这将使更多的临床应用程序能够更快地开发出来，从而推动改善医疗保健。此外，更有效的模式还可以使应用程序在低端硬件上运行，从而使发展中国家能够获得最新的先进临床应用程序。项目的新颖性-当前最先进的深度学习算法的大量数据和功率需求限制了它们的快速部署和广泛使用，这是公认的；降低这些要求仍然是一个热门的研究课题。

Artificial intelligence in the form of deep learning, for instance using convolutional neural networks, has made a huge impact on medical image analysis. It dominates conference and journal publications and has demonstrated state-of-the-art performance in many benchmarks and applications, outperforming human observers in some situations. But, despite this, adoption of these approaches in routine clinical practice has been very slow. One reason for this is that deep learning models are inefficient and expensive to train, often requiring tens or hundreds of thousands of expertly labelled training images, and many days training on high-end GPU hardware. For medical applications the requirement for so much expertly labelled data is a key challenge. After all, a radiologist (a doctor specially trained to interpret medical images) is able to learn new tasks using a far smaller set of training images. This project will investigate approaches to improve the efficiency of training deep learning models, reducing the size and/or level of detail of the required training set whilst maintaining diagnostic accuracy. This would enable more clinical applications to be developed sooner, driving improved healthcare. In addition, more efficient models may also enable applications to run on lower-end hardware, giving developing countries access to the latest advanced clinical applications.Novelty of Project-The extravagant data and power requirements of current state-of-the-art deep learning algorithms that limit their rapid deployment and wide use are well recognized; reducing these requirements remains a hot research topic.