Pancreatic cancer is one of the deadliest types of cancer, with 25% of the diagnosed patients surviving for only one year and 6% of them for five. Computed tomography (CT) screening trials have played a key role in improving early detection of pancreatic cancer, which has shown significant improvement in patient survival rates. However, advanced analysis of such images often requires manual segmentation of the pancreas, which is a time-consuming task. Moreover, pancreas presents high variability in shape, while occupying only a very small area of the entire abdominal CT scans, which increases the complexity of the problem. The rapid development of deep learning can contribute to offering robust algorithms that provide inexpensive, accurate, and user-independent segmentation results that can guide the domain experts. This dissertation addresses this task by investigating a two-step approach for pancreas segmentation, by assisting the task with a prior rough localization or detection of pancreas. This rough localization of the pancreas is provided by an estimated probability map and the detection task is achieved by using the YOLOv4 deep learning algorithm. The segmentation task is tackled by a modified U-Net model applied on cropped data, as well as by using a morphological active contours algorithm. For comparison, the U-Net model was also applied on the full CT images, which provide a coarse pancreas segmentation to serve as reference. Experimental results of the detection network on the National Institutes of Health (NIH) dataset and the pancreas tumour task dataset within the Medical Segmentation Decathlon show 50.67% mean Average Precision. The best segmentation network achieved good segmentation results on the NIH dataset, reaching 67.67% Dice score.

翻译：胰腺癌是最致命的癌症类型之一，确诊患者中仅25%能存活一年，6%能存活五年。计算机断层扫描（CT）筛查试验在改善胰腺癌早期检测方面发挥了关键作用，显著提高了患者生存率。然而，对此类图像的深入分析通常需要手动分割胰腺，这是一项耗时的工作。此外，胰腺形状差异大，且仅占整个腹部CT扫描中极小的区域，进一步增加了问题的复杂性。深度学习的快速发展有助于提供鲁棒算法，这些算法能提供低成本、高精度且不依赖操作者的分割结果，从而指导领域专家。本论文通过研究一种两步法来解决胰腺分割任务：首先通过粗略定位或检测来辅助胰腺分割。该粗略定位通过估计的概率图实现，检测任务则使用YOLOv4深度学习算法完成。分割任务由改进的U-Net模型（应用于裁剪后的数据）以及形态学主动轮廓算法处理。为进行对比，U-Net模型也被应用于完整CT图像，提供粗分割结果作为参考。检测网络在美国国立卫生研究院（NIH）数据集和医学分割十项全能中的胰腺肿瘤任务数据集上取得了50.67%的平均精度均值。最佳分割网络在NIH数据集上取得了良好的分割结果，Dice分数达到67.67%。