Department of Biochemistry and Molecular Biology, Tianjin Medical University Cancer Institute and Hospital, Key Laboratory of Cancer Prevention and Therapy, National Clinical Research Center for Cancer, Tianjin's Clinical Research Center for Cancer, Tianjin 300060, China.
School of Big Data and Software Engineering, Chongqing University, Chongqing 401331, China.
Eur J Radiol. 2022 Sep;154:110443. doi: 10.1016/j.ejrad.2022.110443. Epub 2022 Jul 21.
Accumulated evidence has proven that computer-derived features from computed tomography (CT) through radiomics and deep learning technologies can identify extensive characteristics of pulmonary malignancies, such as nodules detection and malignant lesion discrimination. However, there are few studies on whether CT images can reflect histological subtypes of lung cancer through computer-derived features.
Contrast-enhanced CT images prior treatment from 417 patients diagnosed with small cell lung cancer (SCLC), lung adenocarcinoma (ADC), or lung squamous cell carcinoma (SCC) were collected. ITK-SNAP software was used by trained radiologists for the manual delineation of tumor volume. Patients of each category (SCLC, ADC, SCC) were then randomly split into training datasets and test datasets in an approximately ratio of 8:2. After image pre-processing and augmentation, 25,042 CT images from the training datasets were used to train our self-developed deep learning model for fast-tracking tumor lesions and classifying corresponding histological subtypes simultaneously. The performance of the network was evaluated by accuracy, F1-score and weighted F1-average using 1,921 testing images based on parameters generated during training.
The prediction accuracy of SCLC, ADC, and SCC were 0.83, 0.75 and 0.67, respectively. The weighted F1-average was 0.75. ADC obtained the best F1-score of 0.78, which was outperformed SCLC (0.77) and SCC (0.66). The corresponding AUC values of SCLC, ADC, and SCC were 0.87, 0.84, and 0.76, respectively. Only 0.24 s were required to simultaneously achieve functions of tumor localization and histological classification on a thoracic CT image slice. The heat map visualization illustrated the extracted tumor features to classify subtypes of lung cancer by the proposed model.
The newly developed multi-task algorithm provides a CNN-based DL approach in lung cancer for automatically fast-tracking tumor lesions and classifying corresponding histological subtypes in one-step.
越来越多的证据表明，通过放射组学和深度学习技术从计算机断层扫描（CT）中提取的计算机特征可以识别肺部恶性肿瘤的广泛特征，例如结节检测和恶性病变鉴别。然而，关于 CT 图像是否可以通过计算机衍生特征反映肺癌的组织学亚型，研究甚少。
收集了 417 名经诊断患有小细胞肺癌（SCLC）、肺腺癌（ADC）或肺鳞癌（SCC）的患者治疗前的增强 CT 图像。经过培训的放射科医生使用 ITK-SNAP 软件对肿瘤体积进行手动勾画。然后，将每个类别（SCLC、ADC、SCC）的患者随机分为训练数据集和测试数据集，比例约为 8:2。在图像预处理和扩充后，我们使用来自训练数据集的 25042 张 CT 图像来训练我们自主开发的深度学习模型，以便同时快速跟踪肿瘤病变并对相应的组织学亚型进行分类。使用基于训练过程中生成的参数的 1921 张测试图像，通过准确性、F1 分数和加权 F1 平均值来评估网络的性能。
SCLC、ADC 和 SCC 的预测准确率分别为 0.83、0.75 和 0.67，加权 F1 平均值为 0.75。ADC 获得了最佳的 F1 分数 0.78，优于 SCLC（0.77）和 SCC（0.66）。SCLC、ADC 和 SCC 的相应 AUC 值分别为 0.87、0.84 和 0.76。在一张胸部 CT 图像切片上，同时实现肿瘤定位和组织学分类功能仅需 0.24 秒。热图可视化展示了所提出的模型提取的肿瘤特征，用于对肺癌进行分类。
新开发的多任务算法为肺癌提供了一种基于 CNN 的深度学习方法，可一步实现肿瘤病变的自动快速跟踪和相应组织学亚型的分类。
