Deep learning

Image segmentation

Pattern recognition (psychology)

Mathematics

Biology

Reduction (mathematics)

Engineering

Paleontology

Artificial intelligence

Sensitivity (control systems)

Electronic engineering

Geometry

Nodule (geology)

Computer science

Computer vision

Artificial neural network

False positive paradox

Segmentation

False positive rate

In recent years nodule candidate detection becomes the basis of the automated pulmonary nodule detection system, of which the the upper bound limit performance is determined by the sensitivity of nodule candidates detection. This paper is to improve the nodule candidate detection using deep neural networks. We treat the nodule detection task as pixel-level segmentation problem. Based on the 2D U-NET network. We build a multi-level network to process each CT slice to detect more nodules. Weighted dice loss function is designed to maintain a high sensitivity. More important, different from normaly segmentation problem, it has a heavily unbalanced positive and negative samples. We proposed a training method to make the network converge easily. We further propose an effective non-maximum suppression (NMS) method to remove duplicate nodules. The proposed framework has been validated on LUNA16 dataset. We achieved 94.3% sensitivity score, and had a 1/3 times of false positives less than the official methods of LUNA which is better for false positive reduction task. We provide a deep neural network solution for nodule candidate detection and the experimental result demonstrates the effectiveness of our method. It can also be used for input of the false positive reduction task.

Pulmonary Nodule Detection in CT Images via Deep Neural Network