1. State Key Laboratory of Virtual Reality Technology and Systems, Beihang University, Beijing 100191, China
2. School of Computer Science and Engineering, Nanyang Technological University, Singapore 639798, Singapore
3. Department of Computer Science, Stony Brook University, Stony Brook 11790, USA
Corresponding author (email@example.com
Lung cancer has been the leading cause of cancer-related deaths in 2015 in United States. Early detection of lung nodules will undoubtedly increase the five-year survival rate for lung cancer according to prior studies. In this paper, we propose a novel rating method based on geometrical and statistical features to extract initial nodule candidates and an artificial neural network approach to the detection of lung nodules. The novel method is solely based on 3D distribution of neighboring voxels instead of user-specified features. During initial candidates detection, we combine organized region properties calculated from connected component analysis with corresponding voxel value distributions from statistical analysis to reduce false positives while retaining true nodules. Then we devise multiple artificial neural networks (ANNs) trained from massive voxel neighbor sampling of different types of nodules and organize the outputs using a 3D scoring method to identify final nodules. The experiments on 107 CT cases with 252 nodules in LIDC-IDRI data sets have shown that our new method achieves sensitivity of 89.4% while reducing the false positives to 2.0 per case. Our comprehensive experiments have demonstrated our system would be of great assistance for diagnosis of lung nodules in clinical treatments.
This work was supported in part by National Natural Science Foundation of China (Grant Nos. 61190120, 61190121, 61190125, 61532002, 61300068, 61300067, 61672149, 61672077), National Science Foundation of USA (Grant Nos. IIS-0949467, IIS-1047715, IIS-1049448), Postdoctoral Science Foundation of China (Grant No. 2013M530512), and China Scholarship Council (Grant No. 201506020035).
(Color online) Pipeline of the proposed method. The top row constitutes the initial candidate extraction. The bottom row constitutes the training and forecasting.
(Color online) Architecture of the proposed network. Each network consists of multiple ANNs and all outputs from these ANNs are organized using a 3D scoring method. Finally, a logical AND operator is used to organize all thresholded data and to distinguish true positive and false positive nodules.
(Color online) Pipeline of the potential nodule extraction procedure. (a) Original data. Pixel and slice spacing varies in different data sets. (b) Re-sampled data. The data set is isotropic and pixel spacing as well as slice spacing are interpolated to 1 mm. (c) Extracted lung regions. Juxta-wall nodules are recovered. (d) ROI filtered lung regions. (e) Top: filtered vessels; bottom: thresholded candidates. (f) Final potential nodule candidates.
(Color online) Features used for voting in potential nodule extraction. Both geometrical and statistical features are used.
(Color online) Testing time statistics for all 107 datasets. Maximum, minimum and average time for every nodule at each trained ANN are shown with different colors. The horizonal axis represents indices of the CT cases. The vertical axis represents the corresponding time performance.
(Color online) Statistics for all 107 data sets. The horizonal axes represents the CT case number. The vertical axes represents the corresponding count. We use three colors to label corresponding ground truth, true positive, and false positive nodule counts. It should be noticed that all statistics are at agreement level 3.
(Color online) Four isolated nodules of different sizes with 3D surfaces on the right. (a) and (c) Nodule diameter is 3–5 mm; (b) nodule diameter is over 5 mm; (d) Nodule diameter is over 30 mm. 3D surfaces are enlarged for a better view.
(Color online) Juxta-pleural nodules of different sizes and different conditions with 3D surfaces on the right. (a) and (b) show cases where nodules are slightly attached onto the lung wall; (c) and (d) show cases where nodules are totally attached. Diameters for (a) and (d) are over 30 mm. Diameter for (b) is 5–15 mm. Diameter for protectłinebreak (c) is 3–5 mm. 3D surfaces are enlarged for a better view.
(Color online) Juxta-vessel nodules of different sizes and different conditions. Diameters for (a) and (b) are 5–15 mm. Diameter for (c) is over 30 mm. Only a few vessels are around in (a). (b) is surrounded by tiny vessels. (c) is surrounded by more severe and obvious vessels.
(Color online) Ground-glass optical (GGO) nodules of different sizes and different conditions. Appearances of GGO nodules may vary significantly. (a), (c), and (e) Nodules are slightly attached to lung walls or vessels; (b), (d), and (f) nodules are isolated.
(Color online) Some typical types of false positives not removed by our system. (a) and (c) Nodules are not removed since their appearances and structures are very similar with juxta-vessel nodules. Their locations are stable and not changing along slices. (b) Nodule is not removed since it has a very high intensity inside the vessel which is a direct evidence of being nodules. (d) Nodule is not removed because it is inside a vessel and its size is too small to afford enough structure information.
(Color online) Some typical cases of missed nodules. The nodules are located at the center of the box and the cross hair. (a) Missed nodule is very similar to vessels since its location is changing slice by slice and its distribution is more like an oblique cylinder than a sphere, but it is annotated as nodule by radiologists; (b) very tiny nodule is attached onto lung walls and surrounded by ambiguous voxels, making it hard to find either the correct outline or the distribution. Therefore, our system also refused to classify it as a nodule.
Copyright 2019 Science China Press Co., Ltd. 《中国科学》杂志社有限责任公司 版权所有