[论文解读] Classification of Multiple Diseases on Body CT Scans Using Weakly Supervised Deep Learning- Model weights

Fakrul Islam Tushar, Vincent M. D’Anniballe+6 more|arXiv (Cornell University)|Aug 3, 2020

Radiomics and Machine Learning in Medical Imaging参考文献 33被引用 19

一句话总结

本研究利用从放射科报告中自动提取的标签，开发了用于全身CT扫描多疾病分类的弱监督深度学习模型。通过将基于规则的算法应用于超过13,000份CT检查，该方法在三个器官系统（肺/胸膜、肝脏/胆囊、肾脏/输尿管）的15种疾病标签上实现了0.65至0.97的AUC，展示了无需人工标注即可实现可扩展、高精度的多疾病分类。

ABSTRACT

<p> </p> <h3><strong>Model Documentation: Multidisease Classification Models for Body CT Scans</strong></h3> <p>This document provides an overview and usage guidance for three deep learning models developed to perform multidisease classification on body CT scans. The models are based on 3D convolutional neural networks implemented in <strong>Python using TensorFlow</strong>, and they were trained using weak supervision derived from radiology report text.</p> <h4><strong>Background and Purpose</strong></h4> <p>These models were developed as part of a retrospective study aiming to detect multiple common disease conditions across three major organ systems—lungs and pleura, liver and gallbladder, and kidneys and ureters—using body CT scans. Labels for training were extracted using rule-based natural language processing (NLP) from radiology reports, enabling efficient training without extensive manual annotation.</p> <p>The work demonstrates how weak supervision can support the development of clinically relevant, multi-organ disease classifiers on a large scale.</p> <h4><strong>Model Summary</strong></h4> <p>Each model targets a specific organ system and predicts the presence or absence of five disease categories (four pathologies + one "no apparent disease" class):</p> <ol> <li> <p><strong>Lungs and Pleura: </strong></p> <ul> <li> <p><strong>Labels</strong>: Atelectasis, Nodule, Emphysema, Effusion, No Apparent Disease</p> </li> <li> <p><strong>Performance (AUCs)</strong>:</p> <ul> <li> <p>Atelectasis: 0.77</p> </li> <li> <p>Nodule: 0.65</p> </li> <li> <p>Emphysema: 0.89</p> </li> <li> <p>Effusion: 0.97</p> </li> <li> <p>No Apparent Disease: 0.89</p> </li> </ul> </li> </ul> </li> <li> <p><strong>Liver and Gallbladder</strong></p> <ul> <li> <p><strong>Labels</strong>: Hepatobiliary Calcification, Lesion, Dilation, Fatty Liver, No Apparent Disease</p> </li> <li> <p><strong>Performance (AUCs)</strong>:</p> <ul> <li> <p>Calcification: 0.62</p> </li> <li> <p>Lesion: 0.73</p> </li> <li> <p>Dilation: 0.87</p> </li> <li> <p>Fatty: 0.89</p> </li> <li> <p>No Apparent Disease: 0.82</p> </li> </ul> </li> </ul> </li> <li> <p><strong>Kidneys and Ureters</strong></p> <ul> <li> <p><strong>Labels</strong>: Stone, Atrophy, Lesion, Cyst, No Apparent Disease</p> </li> <li> <p><strong>Performance (AUCs)</strong>:</p> <ul> <li> <p>Stone: 0.83</p> </li> <li> <p>Atrophy: 0.92</p> </li> <li> <p>Lesion: 0.68</p> </li> <li> <p>Cyst: 0.70</p> </li> <li> <p>No Apparent Disease: 0.79</p> </li> </ul> </li> </ul> </li> </ol> <p>The models were trained on CT data from over 13,000 scans and evaluated on a subset of 2,158 volumes with 2,875 manually validated reference labels. Automated label extraction achieved between 91%–99% accuracy during internal validation.</p> <h4><strong>Implementation Details</strong></h4> <ul> <li> <p><strong>Programming Language</strong>: Python</p> </li> <li> <p><strong>Framework</strong>: TensorFlow</p> </li> <li> <p><strong>Model Type</strong>: 3D Convolutional Neural Network (CNN)</p> </li> <li> <p><strong>Preprocessing</strong>: Organ segmentation (via DenseVNet), intensity normalization, and cropping of CT volumes to organ-specific regions of interest.</p> </li> </ul> <h4><strong>Repository Links</strong></h4> <p>The source code, model weights, and usage instructions will be made publicly available through:</p> <ul> <li> <p><strong>GitHub Repository</strong>: https://github.com/fitushar/multi-label-weakly-supervised-classification-of-body-ct</p> </li> <li> <p><strong>GitLab Repository</strong>: https://gitlab.oit.duke.edu/railabs/LoGroup/multi-label-weakly-supervised-classification-of-body-ct</p> </li> </ul> <p>These repositories include:</p> <ul> <li> <p>Model loading and inference scripts</p> </li> <li> <p>Preprocessing pipeline details</p> </li> <li> <p>Instructions for applying the model to new CT data</p> </li> <li> <p>Evaluation tools and AUC reporting scripts</p> </li> </ul> <h4><strong>License and Citation</strong></h4> <p>These models are released for academic research purposes only. If you use them in your work, please cite the original study. Citation details will be provided in the repository README.</p>

研究动机与目标

• 开发用于全身CT的多疾病分类器，以克服人工标注的瓶颈。
• 利用现有放射科报告实现弱监督，减少对昂贵人工标注的依赖。
• 实现在多个器官系统（肺、肝脏、肾脏）中多种疾病类型的多标签分类。
• 验证基于规则的标签提取准确性和在真实世界异构CT数据上的模型性能。
• 证明可扩展、自动化的深度学习流水线在临床影像中的可行性。

提出的方法

• 基于规则的算法从13,667份放射科报告的发现部分提取疾病标签，采用关键词匹配和否定逻辑。
• 标签经人工验证，15种疾病类别中准确率达91–99%。
• 为肺/胸膜、肝脏/胆囊和肾脏/输尿管分别独立训练了三个3D DenseVNet模型。
• 每个模型分类六个标签：每个器官系统五个疾病和“无明显异常”状态。
• 模型使用2 mm × 2 mm × 2 mm各向同性体素，在保持诊断细节的同时降低计算成本。
• 性能通过受试者工作特征（ROC）曲线下面积（AUC）进行评估，使用DeLong方法计算95%置信区间。

实验结果

研究问题

• RQ1 基于规则的放射科报告标签提取能否在全身CT的多种腹部和胸部疾病中实现高准确率？
• RQ2 弱监督的3D卷积神经网络能否在极少人工标注数据下跨多个器官系统实现泛化？
• RQ3 模型性能在全身CT中的局灶性病变与弥漫性病变之间有何差异？
• RQ4 仅使用基于文本的标签，无需解剖定位或人工分割，能否有效实现多疾病分类？
• RQ5 图像分辨率和扫描协议对不同疾病类型模型性能有何影响？

主要发现

• 基于规则的标签提取经人工验证准确率达91–99%，证实了弱监督的可靠性。
• 肺和胸膜模型的AUC分别为：肺不张0.77、结节0.65、肺气肿0.89、胸腔积液0.97、无明显异常0.89。
• 肝脏和胆囊模型的AUC分别为：钙化0.62、病灶0.73、扩张0.87、脂肪变性0.89、无明显异常0.82。
• 肾脏和输尿管模型的AUC分别为：结石0.83、萎缩0.92、病灶0.68、囊肿0.70、无明显异常0.79。
• 弥漫性病变（如肺气肿、脂肪肝）的性能优于局灶性病变（如结节、病灶），尽管肾结石因专用扫描协议而达到高AUC（0.83）。
• 该方法仅使用自由文本报告，实现了在三个器官系统中可扩展、自动化的多疾病分类，显著降低了对人工标注的依赖。