Santone Antonella, Mercaldo Francesco, Brunese Luca
Department of Medicine and Health Sciences "Vincenzo Tiberio", University of Molise, 86100 Campobasso, Italy.
Life (Basel). 2024 Sep 20;14(9):1192. doi: 10.3390/life14091192.
Lung screening is really crucial in the early detection and management of masses, with particular regard to cancer. Studies have shown that lung cancer screening, can reduce lung cancer mortality by 20-30% in high-risk populations. In recent times, the advent of deep learning, with particular regard to computer vision, demonstrated the ability to effectively detect and locate objects from video streams and also (medical) images. Considering these aspects, in this paper, we propose a method aimed to perform instance segmentation, i.e., by providing a mask for each lung mass instance detected, allowing for the identification of individual masses even if they overlap or are close to each other by classifying the detected masses into (generic) nodules, cancer or adenocarcinoma. In this paper, we considered the you-only-look-once model for lung nodule segmentation. An experimental analysis, performed on a set of real-world lung computed tomography images, demonstrated the effectiveness of the proposed method not only in the detection of lung masses but also in lung mass segmentation, thus providing a helpful way not only for radiologist to conduct automatic lung screening but also for discovering very small masses not easily recognizable to the naked eye and that may deserve attention. As a matter of fact, in the evaluation of a dataset composed of 3654 lung scans, the proposed method obtains an average precision of 0.757 and an average recall of 0.738 in the classification task. Additionally, it reaches an average mask precision of 0.75 and an average mask recall of 0.733. These results indicate that the proposed method is capable of not only classifying masses as nodules, cancer, and adenocarcinoma, but also effectively segmenting the areas, thereby performing instance segmentation.
肺部筛查对于肿块的早期检测和管理至关重要，尤其是对于癌症而言。研究表明，肺癌筛查可使高危人群的肺癌死亡率降低20%-30%。近年来，深度学习的出现，特别是计算机视觉方面，展示了从视频流以及（医学）图像中有效检测和定位物体的能力。考虑到这些方面，在本文中，我们提出了一种旨在进行实例分割的方法，即通过为检测到的每个肺部肿块实例提供一个掩码，即使肿块相互重叠或彼此靠近，也能通过将检测到的肿块分类为（一般的）结节、癌症或腺癌来识别单个肿块。在本文中，我们考虑了用于肺结节分割的你只看一次模型。对一组真实世界的肺部计算机断层扫描图像进行的实验分析表明，所提出的方法不仅在肺部肿块检测方面有效，而且在肺部肿块分割方面也有效，从而不仅为放射科医生进行自动肺部筛查提供了一种有用的方法，而且有助于发现肉眼难以识别但可能值得关注的非常小的肿块。事实上，在对由3654次肺部扫描组成的数据集进行评估时，所提出的方法在分类任务中获得了0.757的平均精度和0.738的平均召回率。此外，它的平均掩码精度达到0.75，平均掩码召回率达到0.733。这些结果表明，所提出的方法不仅能够将肿块分类为结节、癌症和腺癌，而且能够有效地分割区域，从而进行实例分割。
