Pulmonary Nodule Detection in Volumetric Chest CT Scans Using CNNs-Based Nodule-Size-Adaptive Detection and Classification

Abstract

In computed tomography, automated detection of pulmonary nodules with a broad spectrum of appearance is still a challenge, especially, in the detection of small nodules. An automated detection system usually contains two major steps: candidate detection and false positive (FP) reduction. We propose a novel strategy for fast candidate detection from volumetric chest CT scans, which can minimize false negatives (FNs) and false positives (FPs). The core of the strategy is a nodule-size-adaptive deep model that can detect nodules of various types, locations, and sizes from 3D images. After candidate detection, each result is located with a bounding cube, which can provide rough size information of the detected objects. Furthermore, we propose a simple yet effective CNNs-based classifier for FP reduction, which benefits from the candidate detection. The performance of the proposed nodule detection was evaluated on both independent and publicly available datasets. Our detection could reach high sensitivity with few FPs and it was comparable with the state-of-the-art systems and manual screenings. The study demonstrated that excellent candidate detection plays an important role in the nodule detection and can simplify the design of the FP reduction. The proposed candidate detection is an independent module, so it can be incorporated with any other FP reduction methods. Besides, it can be used as a potential solution for other similar clinical applications.

Keywords

• Computed tomography
• pulmonary nodule
• object detection
• deep-learning
• convolutional neural networks