Yali Zang

Multi-crop Convolutional Neural Networks for lung nodule malignancy suspiciousness classification

Abstract We investigate the problem of lung nodule malignancy suspiciousness (the likelihood of nodule malignancy) classification using thoracic Computed Tomography (CT) images. Unlike traditional studies primarily relying on cautious nodule segmentation and time-consuming feature extraction, we tackle a more challenging task on directly modeling raw nodule patches and building an end-to-end machine-learning architecture for classifying lung nodule malignancy suspiciousness. We present a Multi-crop Convolutional Neural Network (MC-CNN) to automatically extract nodule salient information by employing a novel multi-crop pooling strategy which crops different regions from convolutional feature maps and then applies max-pooling different times. Extensive experimental results show that the proposed method not only achieves state-of-the-art nodule suspiciousness classification performance, but also effectively characterizes nodule semantic attributes (subtlety and margin) and nodule diameter which are potentially helpful in modeling nodule malignancy.

Multi-scale local binary pattern with filters for spoof fingerprint detection

Fingerprint recognition systems are being increasingly deployed in both government and civilian applications. But the emergence of fake fingerprints brings on a new challenge. Among the numerous fingerprint vitality detection methods, local binary pattern (LBP) is considered as one of the best operators. But the original LBP operator has the limitation of its small spatial support area. So we proposed a novel spoof fingerprint detection method based on multi-scale local binary pattern (MSLBP). Generally speaking, the MSLBP can be realized in two different ways. We implemented both types of MSLBP. Each MSLBP was combined with a set of filters. In this way, each sample in the LBP circle could be made to collect intensity information from a large area rather than a single pixel. The experimental results in the database of the Liveness Detection Competition 2011 (LivDet2011) have shown that both types of MSLBP are effective and superior in spoof fingerprint detection.

A novel ant colony optimization algorithm for large-distorted fingerprint matching

Large distortion may be introduced by non-orthogonal finger pressure and 3D-2D mapping during the process of fingerprint capturing. Furthermore, large variations in resolution and geometric distortion may exist among the fingerprint images acquired from different types of sensors. This distortion greatly challenges the traditional minutiae-based fingerprint matching algorithms. In this paper, we propose a novel ant colony optimization algorithm to establish minutiae correspondences in large-distorted fingerprints. First, minutiae similarity is measured by local features, and an assignment graph is constructed by local search. Then, the minutiae correspondences are established by a pseudo-greedy rule and local propagation, and the pheromone matrix is updated by the local and global update rules. Finally, the minutiae correspondences that maximize the matching score are selected as the matching result. To compensate resolution difference of fingerprint images captured from disparate sensors, a common resolution method is adopted. The proposed method is tested on FVC2004 DB1 and a FINGERPASS cross-matching database established by our lab. The experimental results demonstrate that the proposed algorithm can effectively improve the performance of large-distorted fingerprint matching, especially for those fingerprint images acquired from different modes of acquisition.

Combining features for distorted fingerprint matching

Extracting and fusing discriminative features in fingerprint matching, especially in distorted fingerprint matching, is a challenging task. In this paper, we introduce two novel features to deal with nonlinear distortion in fingerprints. One is finger placement direction which is extracted from fingerprint foreground and the other is ridge compatibility which is determined by the singular values of the affine matrix estimated by some matched minutiae and their associated ridges. Both of them are fixed-length and easy to be incorporated into matching score. In order to improve the matching performance, we combine these two features with orientation descriptor and local minutiae structure, which are used to measure minutiae similarity, to achieve fingerprint matching. In addition, we represent minutiae set as a graph and use graph connect component and iterative robust least square (IRLS) to detect creases and remove spurious minutiae close to creases. Experimental results on FVC2004 DB1 and DB3 demonstrate that the proposed algorithm could obtain promising results. The equal error rates (EER) are 3.35% and 1.49% on DB1 and DB3, respectively.

Multi-scale block local ternary patterns for fingerprints vitality detection

Fingerprint recognition systems are widely deployed in both government and civilian applications. But the emergence of fake fingerprints poses a new threat to privacy security. Among the numerous fingerprint vitality detection methods, the local binary pattern (LBP) is considered as one of the state-of-the-art operators. However, the local binary pattern tends to be sensitive to noise, as there are no types of filters involved in the LBP operator. Worse still, the LBP operator can not reflect these difference whether the pixel value is bigger than the threshold or equal to the threshold. So we proposed a novel fingerprint vitality detection method based on multi-scale block local ternary patterns (MBLTP). Instead of a single pixel, its computation is done based on the average value of blocks. The ternary pattern is adopted to reflect the differences between the pixels and the threshold. The experimental results in the databases of Competition on Fingerprint Liveness Detection 2011 (LivDet 2011) show its superiority.

Learning from Experts: Developing Transferable Deep Features for Patient-Level Lung Cancer Prediction

Due to recent progress in Convolutional Neural Networks (CNNs), developing image-based CNN models for predictive diagnosis is gaining enormous interest. However, to date, insufficient imaging samples with truly pathological-proven labels impede the evaluation of CNN models at scale. In this paper, we formulate a domain-adaptation framework that learns transferable deep features for patient-level lung cancer malignancy prediction. The presented work learns CNN-based features from a large discovery set (2272 lung nodules) with malignancy likelihood labels involving multiple radiologists’ assessments, and then tests the transferable predictability of these CNN-based features on a diagnosis-definite set (115 cases) with true pathologically-proven lung cancer labels. We evaluate our approach on the Lung Image Database Consortium and Image Database Resource Initiative (LIDC-IDRI) dataset, where both human expert labeling information on cancer malignancy likelihood and a set of pathologically-proven malignancy labels were provided. Experimental results demonstrate the superior predictive performance of the transferable deep features on predicting true patient-level lung cancer malignancy (Acc = 70.69 %, AUC = 0.66), which outperforms a nodule-level CNN model (Acc = 65.38 %, AUC = 0.63) and is even comparable to that of using the radiologists’ knowledge (Acc = 72.41 %, AUC = 0.76). The proposed model can largely reduce the demand for pathologically-proven data, holding promise to empower cancer diagnosis by leveraging multi-source CT imaging datasets.

Minutia handedness: A novel global feature for minutiae-based fingerprint matching

Traditional minutiae-based matching algorithms are challenged by the probability that minutiae from different regions of different fingers may not be well matched, and hence lead to erroneous matching results. In this paper we introduce a novel feature called minutia handedness to deal with this problem. First, reference points are detected and additional checking conditions are added to ensure that genuine and accurate reference points can be found. Second, minutia handedness is defined for each minutia according to the bending degree of its associated ridges or the position of the reference points. There are three types of minutiae handedness: right-handed, left-handed and non-handed. Finally, the matching rules between different types of minutiae handedness are set up. The proposed method is tested on eight data sets of FVC2002 (2002) and FVC2004 (2004). The experimental results indicate that the performance of a convectional fingerprint recognition algorithm can be improved by incorporating minutia handedness with a small increment of template size.

Central focused convolutional neural networks: Developing a data-driven model for lung nodule segmentation

HighlightsA data‐driven lung nodule segmentation method without involving shape hypothesis.Two‐branch convolutional neural networks extract both 3D and multi‐scale 2D features.A novel central pooling layer is proposed for feature selection.We propose a weighted sampling method to solve imbalanced training label problem.The method shows strong performance for segmenting juxtapleural nodules. Graphical abstract Figure. No caption available. &NA; Accurate lung nodule segmentation from computed tomography (CT) images is of great importance for image‐driven lung cancer analysis. However, the heterogeneity of lung nodules and the presence of similar visual characteristics between nodules and their surroundings make it difficult for robust nodule segmentation. In this study, we propose a data‐driven model, termed the Central Focused Convolutional Neural Networks (CF‐CNN), to segment lung nodules from heterogeneous CT images. Our approach combines two key insights: 1) the proposed model captures a diverse set of nodule‐sensitive features from both 3‐D and 2‐D CT images simultaneously; 2) when classifying an image voxel, the effects of its neighbor voxels can vary according to their spatial locations. We describe this phenomenon by proposing a novel central pooling layer retaining much information on voxel patch center, followed by a multi‐scale patch learning strategy. Moreover, we design a weighted sampling to facilitate the model training, where training samples are selected according to their degree of segmentation difficulty. The proposed method has been extensively evaluated on the public LIDC dataset including 893 nodules and an independent dataset with 74 nodules from Guangdong General Hospital (GDGH). We showed that CF‐CNN achieved superior segmentation performance with average dice scores of 82.15% and 80.02% for the two datasets respectively. Moreover, we compared our results with the inter‐radiologists consistency on LIDC dataset, showing a difference in average dice score of only 1.98%.

Adaptive Orientation Model Fitting for Latent Overlapped Fingerprints Separation

Overlapped fingerprints are commonly encountered in latent fingerprints lifted from crime scenes. Such overlapped fingerprints can hardly be processed by state-of-the-art fingerprint matchers. Several methods have been proposed to separate the overlapped fingerprints. However, these methods neither provide robust separation results, nor could be generalized for most overlapped fingerprints. In this paper, we propose a novel latent overlapped fingerprints separation algorithm based on adaptive orientation model fitting. Different from existing methods, our algorithm estimates the initial orientation fields in a more accurate way and then separates the orientation fields for component fingerprints through an iterative correction process. Global orientation field models are used to predict and correct the orientations in overlapped regions. Experimental results on the latent overlapped fingerprints database show that the proposed algorithm outperforms the state-of-the-art algorithm in terms of accuracy.

A novel measure of fingerprint image quality using Principal Component Analysis(PCA)

The performance of automatic fingerprint identification system relies heavily on the quality of the fingerprint images. Poor quality images result in missing or spurious features, thus degrading the performance of the identification system. Therefore, it is important for a fingerprint identification system to estimate the quality of the captured fingerprint images. In this paper, a new method based on Principal Component Analysis (PCA) is proposed for fingerprint quality measure. PCA is a common and useful statistical technique for finding patterns in data of high dimension. It can be found that fingerprint patches in a local neighborhood form a simple and regular circular manifold topology in a high-dimensional space. The characterization of manifold topology represents the local properties of the fingerprint. In our method, we first extract two novel features from the expected manifold topology. Then a local block measure of quality is generated according to these two features using multiplication rules. Finally, incorporating the normalized Harris-corner strength (HCS) as weighted value into local block quality measure, we obtain a global quality of a fingerprint image. The proposed method has been evaluated on the databases of fingerprint verification competition 2004DB1 (FVC2004) and our private database(AES2501). The experimental results confirm that the proposed algorithm is simple and effective for fingerprint image quality measure.