More from less: Overcoming Data Scarcity for Deep Learning in Medical Image Computing

该项目旨在通过更有效地利用现有数据，缩小人工智能模型在经典计算机视觉和自然语言处理中的性能与机器学习方法在放射图像中的应用之间的差距。具体而言，我们制定了以下目标：1。利用生成对抗网络（GANs）在3D中更有效地模拟层析医学图像。然后，我们将回答这个问题：在多大程度上，合成数据作为分类问题的额外训练数据是有益的。这将揭示为了在医学图像上以稳定的方式训练GAN需要多少真实数据。此外，这将为人工图像可以在多大程度上提高分类器的性能提供证据。开发归纳迁移学习的模型，可以用更少的数据来训练其他相关问题的算法。模型和权重将作为开源软件发布。这将显示在哪些条件下迁移学习优于从头开始训练。通过开发问题特定的借口任务和损失函数，将自监督学习从自然语言处理（NLP）成功转化为医学图像计算（MIC）问题。这将比传统的监督式学习更有效地利用现有数据。开发概率分割算法，在三维中模拟可能的肿瘤分割分布。这将提供更精确的概率分割，并为许多下游任务（如放射组学）提供基础。

This project aims at closing the gap between the performance of AI models in classic computer vision and natural language processing and the application of machine learning methods to radiological images by leveraging existing data more effectively. Specifically, we formulate the following goals:1. Simulate tomographic medical images using Generative Adversarial Networks (GANs) more effectively by leveraging all spatial context in 3D. We will then answer the question to what degree the synthesized data is beneficial as additional training data for classification problems. This will reveal how much real data is needed in order to train a GAN on medical images in a stable manner. Furthermore, this will provide evidence for the question to what the degree the performance of a classifier may be improved by artificial images.2. Develop models for inductive transfer learning which may be used to train algorithms for other related problems with much less data. The models and weights will be released as open source software. This will show under which conditions transfer learning is superior to training from scratch.3. Translate the success of self-supervised learning from natural language processing (NLP) into medical image computing (MIC) problems by developing problem-specific pretext tasks and loss functions. This will utilize existing data much more efficiently than in tranditional supervised learning.4. Develop probabilistic segmentation algorithms that model the distribution of possible tumor segmentations in 3D. This will provide even more precise probabilistic segmentations and provides a basis for many downstream tasks such as radiomics.