An improved SqueezeNet model for the diagnosis of lung cancer in CT scans
Michail Tsivgoulis, Θωμάς Παπαστεργίου, V. Megalooikonomou
Abstract
Lung cancer is the leading cause of cancer deaths nowadays and its early detection and treatment plays an important role in survival of patients. The main challenge is to acquire an accurate diagnosis in a limited time and without the need of massive computing power. Here, we propose SqueezeNodule-Net, a light and accurate convolutional neural network (CNN) that can rapidly classify nodules into malignant and benign, requiring only a mid-range computing system. It is based on the compact CNN model SqueezeNet and its Fire Module, whose structure we modified in two different ways and compared them with state-of-the-art models. We used 888 CT scans from the public dataset LUNA16 from which, after appropriate preprocessing, we generated 2D 50 × 50 images of benign and malignant nodules. We, also, produced 3D images in order to prove that our models can run successfully with more spatial information by using the same computing system. For 2D images, SqueezeNodule-Net V1 achieves 93.2% accuracy, 94.6% specificity and 89.2% sensitivity, while the SqueezeNodule-Net V2 achieves 94.3% accuracy, 95.3% specificity and 91.3% sensitivity. In 3D space, SqueezeNodule-Net V1 gives 94.3% accuracy, 96.0% specificity and 87.4% sensitivity, while SqueezeNodule-Net V2 gives 95.8% accuracy, 96.2% specificity and 90.2% sensitivity. Overall, compared to Squeeze-Net, SqueezeNodule-Net V1 is 1.2–1.06 times smaller, 1.31–1.5 times faster and has 0.8–2.5 better classification performance, while SqueezeNodule-Net V2 is 1.4–1.5 time larger, 0.04–1.5 times faster and has 0.1–2.7 times better classification performance.