PancreasMRData

Pancreatic assessment through magnetic resonance imaging (MRI) has become increasingly vital for diagnosing anomalies and pathologies, serving as a crucial prerequisite for numerous clinical applications, including diabetes inspection and surgical planning. However, automating pancreas segmentation in medical images remains a formidable challenge. In this paper, we introduce an innovative automated pancreas segmentation approach for MRI images utilizing 2D deep learning. Our methodology employs U-Net, ResU-Net, and nnU-Net architectures, capable of segmenting diverse pancreas samples and identifying pancreas presence/absence in adjacent slices, eliminating the need for manual region of interest or slice selection. Furthermore, we provide a unique and freely available dataset comprising MRI images from 37 patients post-bariatric surgery, greatly enhancing reproducibility and accessibility for AI model development. Our results demonstrate ResU-Net's superior performance with a Dice coefficient of 77.6% (36.55%), and Jaccard index of 69.36% (40.19%). Additionally, we offer a comprehensive critical analysis of the quantification of fatty infiltration, shedding light on potential biases when choosing individual models. Our contribution includes not only a high-quality unique dataset but also an automated approach that simplifies pancreas segmentation, advancing the field's research and clinical applications. Overview Pancreatic MRI images hold a large amount of useful information for medical diagnostics and research. In spite of this, segmenting these images, particularly in order to quantify fat fractions, can be challenging. A complex task is made more challenging by the unique nuances of the pancreas, along with the presence of surrounding fat and exclusions of the pancreatic duct. To advance research in this area, we present the following dataset: 37 Patients: A diverse range of scans that bring forth the variance needed for robust model training. Image Quality Assurance: Every scan of the T1 VIBE DIXON sequence underwent rigorous image quality assessment. Manual Annotations: To save researchers manual labor, the dataset comes with pancreas segmentations, double-checked by seasoned radiologists for accuracy. Motivation This dataset aims to be the bases for researchers looking to pioneer advancements in the area of precise pancreatic fat fraction quantification. Our goal is to provide users with comprehensive, quality-assured collections of images and annotations to enable them to develop innovative solutions, evaluate models, and ultimately improve medical outcomes. Dataset Structure Annotations: Contains annotations for each patient. Data Each patient has a dedicated folder, further divided by the type of MRI sequence. Raw_Dixon_Data: Water and Fat Dixon images, each containing data identified by patient IDs. Preprocessed_dataset: Contains preprocessed data organized by patient IDs.通过磁共振成像（Magnetic Resonance Imaging, MRI）开展胰腺评估，在诊断异常与病理状态中的重要性与日俱增，是包括糖尿病筛查、手术规划在内的诸多临床应用的关键前置条件。然而，医学影像中的胰腺自动化分割仍是一项极具挑战性的任务。本文提出一种创新的MRI影像胰腺自动化分割方法，采用二维深度学习技术。本方法运用U-Net、ResU-Net及nnU-Net架构，可实现多种胰腺样本的分割，并能识别相邻切片中胰腺的存在与否，无需手动选取感兴趣区域或切片。此外，我们还发布了一套独特且免费公开的数据集，包含37例减肥手术后患者的MRI影像，大幅提升了AI模型开发的可复现性与可及性。 本研究结果证实，ResU-Net的性能最优，戴斯系数（Dice coefficient）达77.6%（标准差36.55%），雅卡德指数（Jaccard index）为69.36%（标准差40.19%）。此外，我们还对脂肪浸润量化开展了全面的批判性分析，阐明了选择单一模型时可能存在的偏差。本研究的贡献不仅在于提供了一套高质量的专属数据集，还提出了简化胰腺分割流程的自动化方法，推动了该领域的研究与临床应用进展。 概述 胰腺MRI影像承载着大量可供医疗诊断与研究使用的有效信息。尽管如此，对这些影像进行分割——尤其是为了量化脂肪比例——仍颇具难度。胰腺本身的独特特征、周围脂肪的存在以及胰腺导管的排除因素，进一步加剧了这项复杂任务的难度。为推动该领域的研究，我们发布如下数据集： 37例患者：涵盖多样化的扫描样本，具备训练鲁棒模型所需的各类差异。 影像质量保障：所有T1 VIBE DIXON序列的扫描均经过严格的影像质量评估。 手动标注：为节省研究人员的手动工作量，本数据集附带胰腺分割标注，并由经验丰富的放射科医生双重核查以确保标注准确性。 研究动机 本数据集旨在为致力于精准胰腺脂肪分数量化领域创新研究的科研人员提供基础支撑。我们的目标是为用户提供全面且经过质量验证的影像与标注集，助力其开发创新解决方案、评估模型，并最终改善医疗结局。 数据集结构 标注文件：包含每位患者的标注信息。 数据：每位患者拥有专属文件夹，进一步按MRI序列类型划分。 Raw_Dixon_Data：包含水相与脂肪相Dixon影像，每份数据均以患者ID标识。 Preprocessed_dataset：包含按患者ID组织的预处理数据。