This work achieves a method based on modified extreme learning machine (ELM) with deep convolutional features to detect lung nodules automatically. Convolutional neural networks (CNNs) are employed to extract the features of lung nodules for classification. And then ELM is used to detect the lung nodules by combining the normalization and vote selection. In comparison with the traditional methods, it is shown that our method achieves a higher performance and it can be used as an effective tool for lung nodules computer aided diagnosis.
We propose a novel method for false positive reduction of pulmonary nodules using three-channel samples with different average thickness. A three-channel sample contains a patch centered on the candidate point as well as two patches at the k-th slice above and below the candidate point. Three-channel samples include rich spatial contextual information of pulmonary nodules, and can be trained with a low computational and storage requirement. The convolutional neural networks (CNNs) are constructed and optimized as the feature extractor and classifier of candidates in our study. A fusion method is proposed for fusing multiple prediction results of each candidate. Our method reports high sensitivities of 84.8% and 91.4% at 4 and 8 false positives per scan respectively on 888 CT scans released by the LUNA16 Challenge. The experimental results show that our method significantly reduces false positives in pulmonary nodule detection.
In this paper, we proposed a semi-automatic pulmonary nodule segmentation algorithm, which is operated within a region of interest for each nodule. It mainly includes two parts: the unsupervised training of auto-encoder and the supervised training of segmentation network. Applying an auto-encoder's unsupervised learning, we obtain a feature extractor that consists of its encoded part. Through adding some new neural network layers behind the feature extractor and do supervised learning on it, we get the final segmentation neural network. Compared with the traditional maximum two-dimensional entropy threshold segmentation algorithm, the dice correlation coefficient of this algorithm is 1% - 9% higher in 36 regions of interest segmentation experiments.
Our purposes are to develop a vertebra detection scheme for automated scan planning, which would assist radiological technologists in their routine work for the imaging of vertebrae. Because the orientations of vertebrae were various, and the Haar-like features were only employed to represent the subject on the vertical, horizontal, or diagonal directions, we rotated the CT scout image seven times to make the vertebrae roughly horizontal in least one of the rotated images. Then, we employed Adaboost learning algorithm to construct a strong classifier for the vertebra detection by use of Haar-like features, and combined the detection results with the overlapping region according to the number of times they were detected. Finally, most of the false positives were removed by use of the contextual relationship between them. The detection scheme was evaluated on a database with 76 CT scout image. Our detection scheme reported 1.65 false positives per image at a sensitivity of 94.3% for initial detection of vertebral candidates, and then the performance of detection was improved to 0.95 false positives per image at a sensitivity of 98.6% for the further steps of false positive reduction. The proposed scheme achieved a high performance for the detection of vertebrae with different orientations.