LungNet:一种用于肺肿瘤分类的深度学习框架

LungNet:A deep learning framework for lung tumor classification

肺癌是世界范围内发病率和死亡率最高的恶性肿瘤之一.组织病理切片的检查是病理学家用于评估肺肿瘤类型最可靠的方法.然而深度学习在医学影像分析领域的飞速发展和广泛应用,暗示了放射学数据在进一步描述疾病特征和风险分层方面的有效性.本文基于计算机断层扫描(Computed Tomography,CT)数据,提出了一种LungNet框架用于对良性肺肿瘤、恶性肺肿瘤和感染性病变的分类.该方法将小波池化(Wavelet Pooling)和多通道空间压缩激励(Multi-channel Spatial Squeeze and Excitation,Multi-SE)引入到CNN模型中,利用小波池化代替传统的下采样以减少细节特征的损失,并结合多通道空间压缩激励模块增强各通道之间的依赖性,提高模型学习有效特征的能力.此外,我们在LungNet中融合了放射组学特征,进一步探讨了放射组学特征、深度学习和模型预测结果之间的潜在关系,从而增强模型的可解性.我们在西南医科大学附属医院接受手术治疗的1039名肺癌患者的数据集上训练并测试了LungNet框架,重点关注了良性肺肿瘤和恶性肺肿瘤的分类性能.实验结果表明,LungNet对恶性肺肿瘤识别的AUC达到了0.9795,从而为临床医生提供辅助性诊断.

Lung cancer is one of the malignant tumors with the highest morbidity and mortality worldwide.Tissue sampling for pathologist review is the most reliable method for histology classification,however,the rapid development and widespread application of deep learning in the field of medical image analysis allude to the utility of radiologic data in further describing disease characteristics and for risk stratification.In this study,we propose a LungNet framework for classifying benign lung tumors,malignant lung tumors and infec⁃tious lesions based on computed tomography(CT)data.This method introduces Wavelet Pooling and Multichannel Spatial Squeeze Excitation(Multi-SE)into the CNN model,uses wavelet pooling to replace tradi⁃tional down-sampling to reduce the loss of detailed features,and combines the multi-channel spatial squeeze excitation module to enhance the dependence between channels,improve the model’s ability to learn effective features.In addition,we fusion radiomics features in LungNet to further explore the potential relationship be⁃tween radiomics features,deep learning and model prediction results,thereby enhancing the interpretability of the model.We trained and validated LungNet on a dataset comprising 1039 early-stage lung cancer patients receiving surgical treatment at the Affiliated Hospital of Southwest Medical University,with a focus on the two most common types:Benign Lung Tumors,Malignant Lung Tumors.Deep learning based radiomics can identify histological phenotypes in lung cancer,with AUC of up to 0.9795,and it can provide clinicians with auxiliary diagnosis.