Department of Diagnostic Radiology, Faculty of Applied Medical Sciences, King Abdulaziz University, Jeddah, Saudi Arabia.
Imaging Unit, King Fahd Medical Research Center, King Abdulaziz University, Jeddah, Saudi Arabia.
Microsc Res Tech. 2022 Jun;85(6):2181-2191. doi: 10.1002/jemt.24075. Epub 2022 Feb 4.
Lung's cancer is the leading cause of cancer-related deaths worldwide. Recently cancer mortality rate and incidence increased exponentially. Many patients with lung cancer are diagnosed late, so the survival rate is shallow. Machine learning approaches have been widely used to increase the effectiveness of cancer detection at an early stage. Even while these methods are efficient in detecting specific forms of cancer, there is no known technique that could be used universally and consistently to identify new malignancies. As a result, cancer diagnosis via machine learning algorithms is still fresh area of research. Computed tomography (CT) images are frequently employed for early cancer detection and diagnosis because they contain significant information. In this research, an automated lung cancer detection and classification framework is proposed which consists of preprocessing, three patches local binary pattern feature encoding, local binary pattern, histogram of oriented gradients features are extracted and fused. The fast learning network (FLN) is a novel machine-learning technique that is fast to train and economical in terms of processing resources. However, the FLN's internal power parameters (weight and basis) are randomly initialized, resulting it an unstable algorithm. Therefore, to enhance accuracy, FLN is hybrid with K-nearest neighbors to classify texture and appearance-based features of lung chest CT scans from Kaggle dataset into cancerous and non-cancerous images. The proposed model performance is evaluated using accuracy, sensitivity, specificity on the Kaggle benchmark dataset that is found comparable in state of the art using simple machine learning strategies. RESEARCH HIGHLIGHTS: Fast learning network and K-nearest neighbor hybrid classifier proposed first time for lung cancer classification using handcrafted features including three patches local binary pattern, local binary pattern, and histogram of oriented gradients. Promising results obtained from novel simple combination.
肺癌是全球癌症相关死亡的主要原因。最近，癌症死亡率和发病率呈指数级增长。许多肺癌患者被诊断为时已晚，因此生存率较低。机器学习方法已被广泛用于提高癌症早期检测的效果。尽管这些方法在检测特定类型的癌症方面非常有效，但目前还没有已知的技术可以普遍且一致地识别新的恶性肿瘤。因此，通过机器学习算法进行癌症诊断仍然是一个新兴的研究领域。计算机断层扫描（CT）图像常用于早期癌症检测和诊断，因为它们包含重要信息。在这项研究中，提出了一种自动肺癌检测和分类框架，该框架由预处理、三个补丁局部二值模式特征编码、局部二值模式、方向梯度直方图特征提取和融合组成。快速学习网络（FLN）是一种新型的机器学习技术，它在训练速度和处理资源方面都具有经济性。然而，FLN 的内部功率参数（权重和基）是随机初始化的，因此它是一种不稳定的算法。因此，为了提高准确性，FLN 与 K-最近邻混合用于将 Kaggle 数据集的肺胸部 CT 扫描的纹理和外观特征分类为癌症和非癌症图像。使用 Kaggle 基准数据集评估所提出模型的性能，该数据集在使用简单机器学习策略方面与最新技术具有可比性。研究亮点：首次提出使用快速学习网络和 K-最近邻混合分类器对手工制作的特征（包括三个补丁局部二值模式、局部二值模式和方向梯度直方图）进行肺癌分类。从新颖的简单组合中获得了有希望的结果。
