Segmentation of Lung Nodules on CT Images Using a Nested Three-Dimensional Fully Connected Convolutional Network

• DOI: 10.3389/frai.2022.782225
• Publication Date: 2022-02-17T10:52:27Z
• Authors: Shoji Kido; Shunske Kidera; Yasushi Hirano; Shingo Mabu; Tohru Kamiya; Nobuyuki Tanaka; Yuki Suzuki; Masahiro Yanagawa; Noriyuki Tomiyama
• Keywords: Ground-glass opacity; Nodule (geology); Similarity (geometry)

In computer-aided diagnosis systems for lung cancer, segmentation of nodules is important analyzing image features on computed tomography (CT) images and distinguishing malignant from benign ones. However, it difficult to accurately robustly segment attached the chest wall or with ground-glass opacities using conventional processing methods. Therefore, this study aimed develop a method robust accurate three-dimensional (3D) nodule regions deep learning. study, nested 3D fully connected convolutional network residual unit structures was proposed, designed new loss function. Compared annotated obtained under guidance radiologist, Dice similarity coefficient (DS) intersection over union (IoU) were 0.845 ± 0.008 0.738 0.011, respectively, 332 (lung adenocarcinoma) patients. On other hand, U-Net SegNet, DS 0.822 0.009 0.786 IoU 0.711 0.011 0.660 0.012, respectively. These results indicate that proposed significantly superior well-known learning models. Moreover, we compared those methods, watersheds, graph cuts. The watershed 0.628 0.027 0.494 0.025, cut 0.566 0.025 0.414 0.021, may be useful assist radiologists in such as adenocarcinoma CT images.