Synergistic Representation Learning for Pancreatic Image Analysis

博士项目的目的：在这个项目中，我们的目标是开发用于胰腺图像分割和分析的新型机器学习方法。该项目将实现对胰腺体积和形状的鲁棒和准确的表征，为胰腺解剖结构的评估和病理特征的识别提供定量成像表型。项目描述：胰腺癌是英国第6大最常见的癌症死亡原因，5年生存率仅为5% [1]。然而，如果胰腺癌在早期被诊断出来，手术是可能的，生存率可以高达20% [1]，[2]。胰腺癌的早期诊断具有挑战性，主要是因为症状只发生在晚期，并且仍然缺乏筛查工具。在这个项目中，我们研究了新的机器学习方法，用于从医学图像中自动分割和分析胰腺解剖结构。近年来，人们提出了许多胰腺图像分割的方法。有些是基于图谱的，依赖于图像配准进行图谱传播，然后执行标签融合以创建分割[3]。基于图谱的方法的缺点是，由于多个图像配准的成本，它们在计算上是昂贵的。大多数最近的方法都是基于深度学习的，它们训练卷积神经网络来学习从图像到分割的映射[4]-[10]。由于使用了GPU和一遍推理过程，它们的计算速度更快。最先进的分割方法可以实现正常胰腺的平均Dice重叠度量为86.9%[4]。然而，对于异常胰腺，Dice指标可低至38.4% [4]。这证明了胰腺图像分割中的技术挑战。这些挑战归因于若干因素。首先，胰腺与其他腹部器官相比较小，仅占据3D视场的一小部分。由于类不平衡问题，神经网络对小对象不太敏感。第二，胰腺在解剖形状和外观上高度可变。它的解剖结构随着年龄的增长而改变，导致萎缩、分叶和脂肪变性。对于病理情况，解剖结构也可能受到囊肿和肿瘤的显著影响。第三，神经网络的训练需要大量的数据集。针对临床应用中人工标注的训练数据有限的问题，提出了一种协同表示学习的胰腺图像分割方法，以提高图像分割的鲁棒性和准确性。协同效应将来自多个方面。1)尺度之间的协同作用：多尺度语义信息将以联合和从粗到细的方式合并。2)图像特征和解剖先验之间的协同作用：将学习解剖形状先验以提高分割鲁棒性。3)数据之间的协同作用：完全标记（多器官注释），部分标记（仅胰腺注释）和未注释的数据将用于半监督和部分监督学习。4)模态之间的协同作用：将探索CT和MR模态的语义特征学习。5)计算机与人类之间的协同作用。异常病例和困难的例子将被检测出来，供人类审查和注释，以实现人类在环学习。该项目的输出将是一个自动化工具，可应用于大规模数据集，用于分析胰腺成像表型。预期候选人的背景是工程，计算或物理科学。

Aim of the PhD Project:In this project, we aim to develop novel machine learning approaches for segmentation and analysis of pancreatic images.The project will enable robust and accurate characterisation of pancreatic volumes and shapes, providing quantitative imaging phenotypes for assessment of pancreatic anatomy and identification of pathological features.Project Description:Pancreatic cancer is the 6th most common cause of cancer deaths in the UK with a 5-year survival rate of only 5% [1]. However, if pancreatic cancer is diagnosed at an early stage when surgery is possible, the survival rate can go up to 20% [1], [2]. Early diagnosis of pancreatic cancer is challenging, mainly because symptoms only occur at a late stage and screening tools are still lacking. In this project, we investigate novel machine learning approaches for automated segmentation and analysis of pancreatic anatomy from medical images. It will provide an efficient tool for extracting quantitative image-based biomarkers and assisting clinicians in diagnosis and assessment of pancreatic diseases.A number of methods have been proposed for pancreatic image segmentation in recent years. Some are atlas-based, relying on image registration for atlas propagation and then performing label fusion to create segmentation [3]. A disadvantage with atlas-based methods is that they are computationally expensive due to the cost of multiple image registrations. Most recent methods are deep learning-based, which train convolutional neural networks to learn the mapping from image to segmentation [4]-[10]. They are computationally faster due to the use of GPUs and the one-pass inference process.State-of-the-art segmentation methods can achieve an average Dice overlap metric of 86.9% for normal pancreas [4]. However, for abnormal pancreas, the Dice metric can be as low as 38.4% [4]. This demonstrates the technical challenges in pancreatic image segmentation. The challenges are attributed to several factors. First, the pancreas is small compared to other abdominal organs, occupying only a small proportion of the 3D field-of-view. Neural networks are less sensitive to small objects due to the class imbalance problem. Second, the pancreas is highly variable in anatomical shape and appearance. Its anatomy is altered by ageing, which causes atrophy, lobulation and fatty degeneration. For pathological cases, the anatomy can also be significantly influenced by cysts and tumours. Third, the training of neural networks requires large datasets. Available training data with manual annotations are often limited in clinical scenarios.To address these challenges, we propose a synergistic representation learning approach for pancreatic image segmentation to improve both the robustness and accuracy. The synergy will come from multiple aspects. 1) Synergy between scales: Multi-scale semantic information will be incorporated in a joint and coarse-to-fine fashion. 2) Synergy between image features and anatomical priors: Anatomical shape priors will be learnt to improve segmentation robustness. 3) Synergy between data: Fully-labelled (multi-organ annotation), partially-labelled (pancreas-only annotation) and unannotated data will be utilised for semi- and partially-supervised learning. 4) Synergy between modalities: Both CT and MR modalities will be explored for semantic feature learning. 5) Synergy between computer and human. Abnormal cases and hard examples will be detected for human to review and annotate to enable human-in-the-loop learning.The output of the project will be an automated tool that can be applied to large-scale datasets for analysis of pancreatic imaging phenotypes. The expected candidate's background is engineering, computing or physical sciences.