Attention Networks and Optimized Deep Learning Architectures for Cancer Diagnosis and Prognosis in Medical Imaging

本研究的目标是设计和开发新型深度学习架构，以解决人工智能（AI）在医学领域面临的主要挑战。尽管卷积神经网络（CNN）在计算机视觉方面已经取得了一系列突破，并且在不同的医学成像任务中取得了有希望的结果，例如癌症诊断（例如，肿瘤检测）和预后，存在阻碍其转化为临床环境的功效的未满足的挑战。首先，CNN结果的可视化被认为是将AI整合到临床工作流程中的一个主要挑战，并且在理解图像如何有助于最终结果方面存在知识差距。其次，虽然受试者工作特征曲线下面积（AUC）是医学成像中癌症诊断解决方案的主要评估指标，但CNN不能直接针对AUC进行优化，这可能导致次优结果。第三，作为预处理步骤，大多数基于CNN的诊断和预后解决方案依赖于肿瘤区域（感兴趣区域或ROI）的分割。这通常由临床医生手动完成（例如，放射科医师）或通过使用手动注释训练的分割算法自动或半自动地进行。由于没有明确的方法来确认肿瘤的确切边界，因此ROI注释严重依赖于放射科医师的专业知识和对癌组织的潜在表型及其在医学图像上的外观的理解。这导致由不同放射科医生针对同一病例注释的肿瘤区域的广泛变化，从而导致使用ROI训练的CNN模型的显著变化。感兴趣区域的可变性显着降低了给定标签的基于人工智能的诊断和预后模型的准确性（例如，患者存活率）。在这项研究中，我们将设计、开发和验证深度学习架构，以解决AI在医学成像中的这些主要挑战。我们将为CNN设计和实现三种不同的可视化方法，这些方法可用于CNN可视化和仅使用图像级标签的像素级肿瘤定位。我们将开发一种嵌入CNN架构的遗传算法，使网络能够直接针对AUC进行优化。最后，我们将开发一个深度生成模型，不仅可以发现医学图像和标签之间的关联（例如，癌症等级），它还自动生成图像中驱动这种关联的子区域。我们将提出的解决方案应用于不同的成像方式和癌症部位，包括脑肿瘤（MRI），前列腺癌（MRI）和肺癌（CT）。这项研究的成果将是利用成像数据中潜在的有意义的信息来生成注意力地图的新解决方案，并显着提高CNN在医学成像中癌症诊断和预后方面的性能和可靠性。

The objective of this research is to design and develop novel deep learning architectures to address major challenged that Artificial Intelligence (AI) in Medicine faces. Although Convolutional Neural Networks (CNNs) have shown series of breakthroughs in Computer Vision and have achieved promising results in different Medical Imaging tasks such as cancer diagnosis (e.g., tumour detection) and prognosis, there are unmet challenges that impede their efficacy in translation into clinical settings. First, the visualization of CNNs' results is recognized as a major challenge for the integration of AI into clinical workflow and there is a knowledge gap in understanding how images contribute to the final results. Second, while the Area Under the receiver operating characteristic Curve (AUC) is the main evaluation metric for cancer diagnostic solutions in Medical Imaging, CNNs cannot be directly optimized for AUC, which may lead to suboptimal results. Third, as a preprocessing step, most CNN-based diagnostic and prognostic solutions rely on segmentation of tumour regions (region of interest or ROI). This is usually done manually by a clinician (e.g., radiologist) or automatically or semi-automatically by a segmentation algorithm, which is trained using the manual annotations. Because there is no definite way to confirm the exact boundaries of a tumour, the ROI annotation therefore heavily relies on radiologists' expertise and understanding of the underlying phenotype of the cancerous tissue and its appearance on the medical images. This leads to a wide variation of tumour regions annotated by different radiologists for the same case resulting in a significant variation in CNN models trained using ROIs. ROI variability significantly decreases the accuracy of AI-based diagnostic and prognostic models for a given label (e.g., patient survival). In this research, we will design, develop, and validate deep learning architectures that address these major challenges for AI in Medical Imaging. We will design and implement three different visualization methods for CNNs, which can be used for both CNN visualization and tumour localization at pixel level using image-level labels only. We will develop a genetic algorithm embedded into CNN architecture which enables the network to be directly optimized for AUC. Finally, we will develop a deep generative model that not only discovers the associations between medical images and the label (e.g., cancer grade), it also automatically generates the subregions in the image which drive such associations. We will apply the proposed solutions to different imaging modalities and cancer sites including brain tumours (MRI), prostate cancer (MRI), and lung cancer (CT). The outcome of this research will be novel solutions to harness meaningful information latent in imaging data to generate attention maps and significantly improve the performance and reliability of CNNs for both cancer diagnosis and prognosis in Medical Imaging.