使用 3-D 卷積神經網路於肺部 CT 結節診斷

Abstract

肺癌是全球癌症發生和死亡的主要原因之一。此外，肺癌的預後仍然不理想,大部分國家的五年存活率低於20%。而低劑量電腦斷層掃描(low-dose computed tomography, LDCT)是一種廣泛使用的三維(three-dimensional, 3-D)篩檢方法，可以幫助醫生檢查肺結節，從而降低死亡率。實體結節(solid nodule, SN)、毛玻璃樣陰影(ground-glass opacity, GGO)和部分實體結節(part-solid nodule, PSN)是三種具有不同密度和良惡性程度的結節類型。準確的結節分類是一項具有挑戰性的工作，但可以通過基於深度學習的電腦輔助診斷(computer-aided diagnosis, CADx)系統得到改善。卷積神經網絡(convolutional neural network, CNN)和注意力機制現在是電腦輔助診斷系統設計中最常用的深度學習方法，因為它們具有強大的特徵提取和重新加權能力。因此，本研究提出了一種基於卷積神經網絡架構和注意力機制的三維電腦輔助診斷系統，用於結節類型分類。我們的電腦輔助診斷系統包括影像前處理和結節分類兩個部分。首先進行影像前處理，包括感興趣區域(volumes of interest, VOI)提取和影像縮放，以從低劑量電腦斷層掃描影像中提取結節及其周圍組織。然後，將提取的感興趣區域輸入到結節分類模型中，也就是我們提出的三維SE-Inception模型，它由Inception-v4、Inception-ResNet-v2和squeeze-and-excitation(SE)模組構建而成，可以預測結節類型。此外，我們也提出F1引導的動態平衡訓練(F1-guided dynamic balance training, FDBT)方法，以實現更快的訓練和更好的性能。進行實驗時，從8,789個低劑量電腦斷層掃描影像提取的34,898個肺結節被使用於評估系統。根據實驗結果， 我們的電腦輔助診斷系統達到90.0%的micro accuracy和83.6%的macro F1-score，這證明了其結節分類的有效性。

Lung cancer is a leading cause of cancer incidence and mortality worldwide. The prognosis for lung cancer remains unsatisfactory, with five-year survival rates below 20% in most countries. Low-dose computed tomography (LDCT) is a widely used three-dimensional (3-D) screening modality to help physicians check lung nodules to reduce mortality. Solid nodules (SNs), ground-glass opacities (GGOs), and part-solid nodules (PSNs) are three types of nodules with different densities and degrees of benignity and malignancy. Accurate nodule classification is challenging but could be improved with a computer-aided diagnosis (CADx) system based on deep learning. Convolutional neural networks (CNN) and attention mechanisms are now two of the most commonly used deep learning methods for CADx system design due to the powerful feature extraction and reweighting. Therefore, a 3-D CADx based on CNN architecture and attention scheme is proposed for nodule-type classification. Our CADx system consists of image preprocessing and nodule classification. First, the image preprocessing composed of volume of interest (VOI) extraction and image resizing is performed to crop nodules and surrounding tissues from LDCT images. Then, the extracted VOIs are fed into the nodule classification model. Our classification model, the 3-D SE-Inception model, is built with Inception-v4, Inception-ResNet-v2, and squeeze-and-excitation (SE) modules to determine nodule types. Furthermore, the F1-guided dynamic balance training (FDBT) is introduced for faster training and better performance. In experiments, 34,898 pulmonary nodules from 8,789 3-D lung CT scans were used for system evaluation. According to the experiment, the CADx system achieved 90.0% micro accuracy and 83.6% macro F1-score, which proved its nodule classification effectiveness.