Enabling Visual Object Detection With Object Sounds via Visual Modality Recalling Memory.

When humans hear the sound of an object, they recall associated visual information and integrate the sound with recalled visual modality to detect the object. In this article, we present a novel sound-based object detector that mimics this process. We design a visual modality recalling (VMR) memory to recall information of a visual modality based on an audio modal input (i.e., sound). To achieve this goal, we propose a VMR loss and an audio-visual association loss to guide the VMR memory to memorize visual modal information by establishing associations between audio and visual modalities. With the visual modal information recalled through the VMR memory along with the original audio input, we perform audio-visual integration. In this step, we introduce an integrated feature contrastive loss that allows the integrated feature to be embedded as if it were encoded using both audio and visual modal inputs. This guidance enables our sound-based object detector to effectively perform visual object detection even when only sound is provided. We believe that our work is a cornerstone study that offers a new perspective to conventional object detection studies that solely rely on the visual modality. Comprehensive experimental results demonstrate the effectiveness of the proposed method with the VMR memory.

Deep learning-based classification system of bacterial keratitis and fungal keratitis using anterior segment images.

Introduction: Infectious keratitis is a vision threatening disease. Bacterial and fungal keratitis are often confused in the early stages, so right diagnosis and optimized treatment for causative organisms is crucial. Antibacterial and antifungal medications are completely different, and the prognosis for fungal keratitis is even much worse. Since the identification of microorganisms takes a long time, empirical treatment must be started according to the appearance of the lesion before an accurate diagnosis. Thus, we developed an automated deep learning (DL) based diagnostic system of bacterial and fungal keratitis based on the anterior segment photographs using two proposed modules, Lesion Guiding Module (LGM) and Mask Adjusting Module (MAM). Methods: We used 684 anterior segment photographs from 107 patients confirmed as bacterial or fungal keratitis by corneal scraping culture. Both broad- and slit-beam images were included in the analysis. We set baseline classifier as ResNet-50. The LGM was designed to learn the location information of lesions annotated by ophthalmologists and the slit-beam MAM was applied to extract the correct feature points from two different images (broad- and slit-beam) during the training phase. Our algorithm was then externally validated using 98 images from Google image search and ophthalmology textbooks. Results: A total of 594 images from 88 patients were used for training, and 90 images from 19 patients were used for test. Compared to the diagnostic accuracy of baseline network ResNet-50, the proposed method with LGM and MAM showed significantly higher accuracy (81.1 vs. 87.8%). We further observed that the model achieved significant improvement on diagnostic performance using open-source dataset (64.2 vs. 71.4%). LGM and MAM module showed positive effect on an ablation study. Discussion: This study demonstrated that the potential of a novel DL based diagnostic algorithm for bacterial and fungal keratitis using two types of anterior segment photographs. The proposed network containing LGM and slit-beam MAM is robust in improving the diagnostic accuracy and overcoming the limitations of small training data and multi type of images.

Stereoscopic Vision Recalling Memory for Monocular 3D Object Detection.

Monocular 3D object detection has drawn increasing attention in various human-related applications, such as autonomous vehicles, due to its cost-effective property. On the other hand, a monocular image alone inherently contains insufficient information to infer the 3D information. In this paper, we propose a new monocular 3D object detector that can recall the stereoscopic visual information about an object, given a left-view monocular image. Here, we devise a location embedding module to handle each object by being aware of its location. Next, given the object appearance of the left-view monocular image, we devise Monocular-to-Stereoscopic (M2S) memory that can recall the object appearance of the right-view and depth information. For this purpose, we introduce a stereoscopic vision memorizing loss that guides the M2S memory to store the stereoscopic visual information. Furthermore, we propose a binocular vision association loss to guide the M2S memory that can associate the information of the left-right view about the object when estimating the depth. As a result, our monocular 3D object detector with the M2S memory can effectively exploit the recalled stereoscopic visual information in the inference phase. The comprehensive experimental results on two public datasets, KITTI 3D Object Detection Benchmark and Waymo Open Dataset, demonstrate the effectiveness of the proposed method. We claim that our method is a step-forward method that follows the behaviors of humans that can recall the stereoscopic visual information even when one eye is closed.

Advancing Adversarial Training by Injecting Booster Signal.

Recent works have demonstrated that deep neural networks (DNNs) are highly vulnerable to adversarial attacks. To defend against adversarial attacks, many defense strategies have been proposed, among which adversarial training (AT) has been demonstrated to be the most effective strategy. However, it has been known that AT sometimes hurts natural accuracy. Then, many works focus on optimizing model parameters to handle the problem. Different from the previous approaches, in this article, we propose a new approach to improve the adversarial robustness using an external signal rather than model parameters. In the proposed method, a well-optimized universal external signal called a booster signal is injected into the outside of the image which does not overlap with the original content. Then, it boosts both adversarial robustness and natural accuracy. The booster signal is optimized in parallel to model parameters step by step collaboratively. Experimental results show that the booster signal can improve both the natural and robust accuracies over the recent state-of-the-art AT methods. Also, optimizing the booster signal is general and flexible enough to be adopted on any existing AT methods.

Defending Person Detection Against Adversarial Patch Attack by Using Universal Defensive Frame.

Person detection has attracted great attention in the computer vision area and is an imperative element in human-centric computer vision. Although the predictive performances of person detection networks have been improved dramatically, they are vulnerable to adversarial patch attacks. Changing the pixels in a restricted region can easily fool the person detection network in safety-critical applications such as autonomous driving and security systems. Despite the necessity of countering adversarial patch attacks, very few efforts have been dedicated to defending person detection against adversarial patch attack. In this paper, we propose a novel defense strategy that defends against an adversarial patch attack by optimizing a defensive frame for person detection. The defensive frame alleviates the effect of the adversarial patch while maintaining person detection performance with clean person. The proposed defensive frame in the person detection is generated with a competitive learning algorithm which makes an iterative competition between detection threatening module and detection shielding module in person detection. Comprehensive experimental results demonstrate that the proposed method effectively defends person detection against adversarial patch attacks.

Robust Perturbation for Visual Explanation: Cross-Checking Mask Optimization to Avoid Class Distortion.

Along with the outstanding performance of the deep neural networks (DNNs), considerable research efforts have been devoted to finding ways to understand the decision of DNNs structures. In the computer vision domain, visualizing the attribution map is one of the most intuitive and understandable ways to achieve human-level interpretation. Among them, perturbation-based visualization can explain the "black box" property of the given network by optimizing perturbation masks that alter the network prediction of the target class the most. However, existing perturbation methods could make unexpected changes to network predictions after applying a perturbation mask to the input image, resulting in a loss of robustness and fidelity of the perturbation mechanisms. In this paper, we define class distortion as the unexpected changes of the network prediction during the perturbation process. To handle that, we propose a novel visual interpretation framework, Robust Perturbation, which shows robustness against the unexpected class distortion during the mask optimization. With a new cross-checking mask optimization strategy, our proposed framework perturbs the target prediction of the network while upholding the non-target predictions, providing more reliable and accurate visual explanations. We evaluate our framework on three different public datasets through extensive experiments. Furthermore, we propose a new metric for class distortion evaluation. In both quantitative and qualitative experiments, tackling the class distortion problem turns out to enhance the quality and fidelity of the visual explanation in comparison with the existing perturbation-based methods.

BMAN: Bidirectional Multi-scale Aggregation Networks for Abnormal Event Detection.

Abnormal event detection is an important task in video surveillance systems. In this paper, we propose a novel bidirectional multi-scale aggregation networks (BMAN) for abnormal event detection. The proposed BMAN learns spatiotemporal patterns of normal events to detect deviations from the learned normal patterns as abnormalities. The BMAN consists of two main parts: an inter-frame predictor and an appearancemotion joint detector. The inter-frame predictor is devised to encode normal patterns, which generates an inter-frame using bidirectional multi-scale aggregation based on attention. With the feature aggregation, robustness for object scale variations and complex motions is achieved in normal pattern encoding. Based on the encoded normal patterns, abnormal events are detected by the appearance-motion joint detector in which both appearance and motion characteristics of scenes are considered. Comprehensive experiments are performed, and the results show that the proposed method outperforms the existing state-of-the-art methods. The resulting abnormal event detection is interpretable on the visual basis of where the detected events occur. Further, we validate the effectiveness of the proposed network designs by conducting ablation study and feature visualization.

Endometrium segmentation on transvaginal ultrasound image using key-point discriminator.

PURPOSE: Transvaginal ultrasound imaging provides useful information for diagnosing endometrial pathologies and reproductive health. Endometrium segmentation in transvaginal ultrasound (TVUS) images is very challenging due to ambiguous boundaries and heterogeneous textures. In this study, we developed a new segmentation framework which provides robust segmentation against ambiguous boundaries and heterogeneous textures of TVUS images. METHODS: To achieve endometrium segmentation from TVUS images, we propose a new segmentation framework with a discriminator guided by four key points of the endometrium (namely, the endometrium cavity tip, the internal os of the cervix, and the two thickest points between the two basal layers on the anterior and posterior uterine walls). The key points of the endometrium are defined as meaningful points that are related to the characteristics of the endometrial morphology, namely the length and thickness of the endometrium. In the proposed segmentation framework, the key-point discriminator distinguishes a predicted segmentation map from a ground-truth segmentation map according to the key-point maps. Meanwhile, the endometrium segmentation network predicts accurate segmentation results that the key-point discriminator cannot discriminate. In this adversarial way, the key-point information containing endometrial morphology characteristics is effectively incorporated in the segmentation network. The segmentation network can accurately find the segmentation boundary while the key-point discriminator learns the shape distribution of the endometrium. Moreover, the endometrium segmentation can be robust to the heterogeneous texture of the endometrium. We conducted an experiment on a TVUS dataset that contained 3,372 sagittal TVUS images and the corresponding key points. The dataset was collected by three hospitals (Ewha Woman's University School of Medicine, Asan Medical Center, and Yonsei University College of Medicine) with the approval of the three hospitals' Institutional Review Board. For verification, fivefold cross-validation was performed. RESULT: The proposed key-point discriminator improved the performance of the endometrium segmentation, achieving 82.67 % for the Dice coefficient and 70.46% for the Jaccard coefficient. In comparison, on the TVUS images UNet, showed 58.69 % for the Dice coefficient and 41.59 % for the Jaccard coefficient. The qualitative performance of the endometrium segmentation was also improved over the conventional deep learning segmentation networks. Our experimental results indicated robust segmentation by the proposed method on TVUS images with heterogeneous texture and unclear boundary. In addition, the effect of the key-point discriminator was verified by an ablation study. CONCLUSION: We proposed a key-point discriminator to train a segmentation network for robust segmentation of the endometrium with TVUS images. By utilizing the key-point information, the proposed method showed more reliable and accurate segmentation performance and outperformed the conventional segmentation networks both in qualitative and quantitative comparisons.

VRSA Net: VR Sickness Assessment Considering Exceptional Motion for 360° VR Video.

The viewing safety is one of the main issues in viewing virtual reality (VR) content. In particular, VR sickness could occur when watching immersive VR content. To deal with the viewing safety for VR content, objective assessment of VR sickness is of great importance. In this paper, we propose a novel objective VR sickness assessment (VRSA) network based on deep generative model for automatically predicting the VR sickness score. The proposed method takes into account motion patterns of VR videos in which an exceptional motion is a critical factor inducing excessive VR sickness in human motion perception. The proposed VRSA network consists of two parts, which are VR video generator and VR sickness score predictor. By training the VR video generator with common videos with non-exceptional motion, the generator learns the tolerance of VR sickness in human motion perception. As a result, the difference between the original and the generated videos by the VR video generator could represent exceptional motion of VR video causing VR sickness. In the VR sickness score predictor, the VR sickness score is predicted by projecting the difference between the original and the generated videos onto the subjective score space. For the evaluation of VR sickness assessment, we built a new dataset which consists of 360° videos (stimuli), corresponding physiological signals, and subjective questionnaires from subjective assessment experiments. Experimental results demonstrated that the proposed VRSA network achieved a high correlation with human perceptual score for VR sickness.

Visually interpretable deep network for diagnosis of breast masses on mammograms.

Recently, deep learning technology has achieved various successes in medical image analysis studies including computer-aided diagnosis (CADx). However, current CADx approaches based on deep learning have a limitation in interpreting diagnostic decisions. The limited interpretability is a major challenge for practical use of current deep learning approaches. In this paper, a novel visually interpretable deep network framework is proposed to provide diagnostic decisions with visual interpretation. The proposed method is motivated by the fact that the radiologists characterize breast masses according to the breast imaging reporting and data system (BIRADS). The proposed deep network framework consists of a BIRADS guided diagnosis network and a BIRADS critic network. A 2D map, named BIRADS guide map, is generated in the inference process of the deep network. The visual features extracted from the breast masses could be refined by the BIRADS guide map, which helps the deep network to focus on more informative areas. The BIRADS critic network makes the BIRADS guide map to be relevant to the characterization of masses in terms of BIRADS description. To verify the proposed method, comparative experiments have been conducted on public mammogram database. On the independent test set (170 malignant masses and 170 benign masses), the proposed method was found to have significantly higher performance compared to the deep network approach without using the BIRADS guide map (p < 0.05). Moreover, the visualization was conducted to show the location where the deep network exploited more information. This study demonstrated that the proposed visually interpretable CADx framework could be a promising approach for visually interpreting the diagnostic decision of the deep network.

Latent feature representation with depth directional long-term recurrent learning for breast masses in digital breast tomosynthesis.

Characterization of masses in computer-aided detection systems for digital breast tomosynthesis (DBT) is an important step to reduce false positive (FP) rates. To effectively differentiate masses from FPs in DBT, discriminative mass feature representation is required. In this paper, we propose a new latent feature representation boosted by depth directional long-term recurrent learning for characterizing malignant masses. The proposed network is designed to encode mass characteristics in two parts. First, 2D spatial image characteristics of DBT slices are encoded as a slice feature representation by convolutional neural network (CNN). Then, depth directional characteristics of masses among the slice feature representations are encoded by the proposed depth directional long-term recurrent learning. In addition, to further improve the class discriminability of latent feature representation, we have devised three objective functions aiming to (a) minimize classification error, (b) minimize intra-class variation within the same class, and (c) preserve feature representation consistency in a central slice. Experimental results have demonstrated that the proposed latent feature representation achieves a higher level of classification performance in terms of receiver operating characteristic (ROC) curves and the area under the ROC curve values compared to performance with feature representation learned by conventional CNN and hand-crafted features.

Detection of masses in digital breast tomosynthesis using complementary information of simulated projection.

PURPOSE: The purpose of this study is to develop a computer-aided detection system that combines the detection results in 3D digital breast tomosynthesis (DBT) volume and 2D simulated projection (synthesized image which is not provided by the vendor but generated from DBT volume in this study) to improve the accuracy of mass detection in DBT. METHODS: The 3D DBT volume has a problem of blurring in the out-of-focus plane because it is reconstructed from a limited number of projection view images acquired over a limited angular range. To solve the problem, the simulated projection is generated by measuring the blurriness of voxels in the DBT volume and adopting conspicuity voxels. A contour-based detection algorithm is applied to detecting masses in the simulated projection. The DBT volume is analyzed by using an unsupervised mass detection algorithm, which results in mass candidates in the DBT volume. The mass likelihood scores estimated for mass candidates on the DBT volume and the simulated projection are merged in a probabilistic manner through a Bayesian network model to differentiate masses and false positives (FPs). Experiments were conducted on a clinical data set of 320 DBT volumes. In 90 volumes, at least one biopsy-proven malignant mass was presented. The longest diameter of masses ranged from 7.0 to 56.4 mm (mean = 25.4 mm). The sizes of masses in the data set were relatively large compared to the sizes of the masses reported in other detection studies. Three image quality measurements (overall sharpness, sharpness of mass boundary, and contrast) were used to evaluate the image quality of the simulated projection compared to the DBT central slice where the mass was most conspicuous and other projection methods (maximum intensity projection and average projection). A free-response receiver operating characteristic (FROC) analysis was adopted for evaluating the accuracy of mass detection in the DBT volume, the simulated projection, and the combined approach. A jackknife FROC analysis (JAFROC) was used to estimate the statistical significance of the difference between two FROC curves. RESULTS: The overall sharpness and the sharpness of mass boundary in the simulated projection are higher than those in the DBT central slice and other projection methods. The contrast of the simulated projection is lower than the DBT central slice. The mass detection in the DBT volume achieved region-based sensitivities of 80% and 85% with 1.75 and 2.11 FPs per DBT volume. The proposed combined mass detection approach achieved same sensitivities with reduced FPs of 1.33 and 1.93 per DBT volume. The difference of the FROC curves between the combined approach and the mass detection in the DBT volume was statistically significant (p < 0.01) by JAFROC analysis. CONCLUSIONS: This study indicates that the combined approach that merges the detection results in the DBT volume and the simulated projection is a promising approach to improve the accuracy of mass detection in DBT.

Improving mass detection using combined feature representations from projection views and reconstructed volume of DBT and boosting based classification with feature selection.

In digital breast tomosynthesis (DBT), image characteristics of projection views and reconstructed volume are different and both have the advantage of detecting breast masses, e.g. reconstructed volume mitigates a tissue overlap, while projection views have less reconstruction blur artifacts. In this paper, an improved mass detection is proposed by using combined feature representations from projection views and reconstructed volume in the DBT. To take advantage of complementary effects on different image characteristics of both data, combined feature representations are extracted from both projection views and reconstructed volume concurrently. An indirect region-of-interest segmentation in projection views, which projects volume-of-interest in reconstructed volume into the corresponding projection views, is proposed to extract combined feature representations. In addition, a boosting based classification with feature selection has been employed for selecting effective feature representations among a large number of combined feature representations, and for reducing false positives. Experiments have been conducted on a clinical data set that contains malignant masses. Experimental results demonstrate that the proposed mass detection can achieve high sensitivity with a small number of false positives. In addition, the experimental results demonstrate that the selected feature representations for classifying masses complementarily come from both projection views and reconstructed volume.

Region based stellate features combined with variable selection using AdaBoost learning in mammographic computer-aided detection.

In this paper, a new method is developed for extracting so-called region-based stellate features to correctly differentiate spiculated malignant masses from normal tissues on mammograms. In the proposed method, a given region of interest (ROI) for feature extraction is divided into three individual subregions, namely core, inner, and outer parts. The proposed region-based stellate features are then extracted to encode the different and complementary stellate pattern information by computing the statistical characteristics for each of the three different subregions. To further maximize classification performance, a novel variable selection algorithm based on AdaBoost learning is incorporated for choosing an optimal subset of variables of region-based stellate features. In particular, we develop a new variable selection metric (criteria) that effectively determines variable importance (ranking) within the conventional AdaBoost framework. Extensive and comparative experiments have been performed on the popular benchmark mammogram database (DB). Results show that our region-based stellate features (extracted from automatically segmented ROIs) considerably outperform other state-of-the-art features developed for mammographic spiculated mass detection or classification. Our results also indicate that combining region-based stellate features with the proposed variable selection strategy has an impressive effect on improving spiculated mass classification and detection.

Breast mass detection using slice conspicuity in 3D reconstructed digital breast volumes.

In digital breast tomosynthesis, the three dimensional (3D) reconstructed volumes only provide quasi-3D structure information with limited resolution along the depth direction due to insufficient sampling in depth direction and the limited angular range. The limitation could seriously hamper the conventional 3D image analysis techniques for detecting masses because the limited number of projection views causes blurring in the out-of-focus planes. In this paper, we propose a novel mass detection approach using slice conspicuity in the 3D reconstructed digital breast volumes to overcome the above limitation. First, to overcome the limited resolution along the depth direction, we detect regions of interest (ROIs) on each reconstructed slice and separately utilize the depth directional information to combine the ROIs effectively. Furthermore, we measure the blurriness of each slice for resolving the degradation of performance caused by the blur in the out-of-focus plane. Finally, mass features are extracted from the selected in focus slices and analyzed by a support vector machine classifier to reduce the false positives. Comparative experiments have been conducted on a clinical data set. Experimental results demonstrate that the proposed approach outperforms the conventional 3D approach by achieving a high sensitivity with a small number of false positives.

Computer-aided detection (CAD) of breast masses in mammography: combined detection and ensemble classification.

We propose a novel computer-aided detection (CAD) framework of breast masses in mammography. To increase detection sensitivity for various types of mammographic masses, we propose the combined use of different detection algorithms. In particular, we develop a region-of-interest combination mechanism that integrates detection information gained from unsupervised and supervised detection algorithms. Also, to significantly reduce the number of false-positive (FP) detections, the new ensemble classification algorithm is developed. Extensive experiments have been conducted on a benchmark mammogram database. Results show that our combined detection approach can considerably improve the detection sensitivity with a small loss of FP rate, compared to representative detection algorithms previously developed for mammographic CAD systems. The proposed ensemble classification solution also has a dramatic impact on the reduction of FP detections; as much as 70% (from 15 to 4.5 per image) at only cost of 4.6% sensitivity loss (from 90.0% to 85.4%). Moreover, our proposed CAD method performs as well or better (70.7% and 80.0% per 1.5 and 3.5 FPs per image respectively) than the results of mammography CAD algorithms previously reported in the literature.

Denoising 3D ultrasound volumes using sparse representation.

In this paper, a new 3D ultrasound (US) denoising technique that adopts the sparse representation has been proposed for an effective noise reduction in 3D US volumes. The purpose of the proposed method is to reduce image noise while preserving 3D objects edges, hence improving the human interpretation for clinical diagnosis and the 3D segmentation accuracy for further automatic malignancy detection. For denoising 3D US volumes, sparse representation was employed, which has showed an excellent performance in reducing Gaussian noise. It has been well known that US images contain severe multiplicative speckle noise, which has different characteristics compared to the additive Gaussian noise. In this paper, we propose a denoising framework for effectively reducing both Gaussian noise and speckle noise on 3D US volumes. The proposed method removes Gaussian noise using sparse representation. Then, a logarithmic transform is performed to transform the speckle noise into Gaussian noise for applying the sparse representation. To demonstrate the effectiveness of the proposed denoising method, comparative and quantitative experiments had been conducted on a synthesized 3D US phantom data. Experimental results showed that the proposed denoising could improve image quality in terms of denoising measurements.

Mass type-specific sparse representation for mass classification in computer-aided detection on mammograms.

BACKGROUND: Breast cancer is the leading cause of both incidence and mortality in women population. For this reason, much research effort has been devoted to develop Computer-Aided Detection (CAD) systems for early detection of the breast cancers on mammograms. In this paper, we propose a new and novel dictionary configuration underpinning sparse representation based classification (SRC). The key idea of the proposed algorithm is to improve the sparsity in terms of mass margins for the purpose of improving classification performance in CAD systems. METHODS: The aim of the proposed SRC framework is to construct separate dictionaries according to the types of mass margins. The underlying idea behind our method is that the separated dictionaries can enhance the sparsity of mass class (true-positive), leading to an improved performance for differentiating mammographic masses from normal tissues (false-positive). When a mass sample is given for classification, the sparse solutions based on corresponding dictionaries are separately solved and combined at score level. Experiments have been performed on both database (DB) named as Digital Database for Screening Mammography (DDSM) and clinical Full Field Digital Mammogram (FFDM) DBs. In our experiments, sparsity concentration in the true class (SCTC) and area under the Receiver operating characteristic (ROC) curve (AUC) were measured for the comparison between the proposed method and a conventional single dictionary based approach. In addition, a support vector machine (SVM) was used for comparing our method with state-of-the-arts classifier extensively used for mass classification. RESULTS: Comparing with the conventional single dictionary configuration, the proposed approach is able to improve SCTC of up to 13.9% and 23.6% on DDSM and FFDM DBs, respectively. Moreover, the proposed method is able to improve AUC with 8.2% and 22.1% on DDSM and FFDM DBs, respectively. Comparing to SVM classifier, the proposed method improves AUC with 2.9% and 11.6% on DDSM and FFDM DBs, respectively. CONCLUSIONS: The proposed dictionary configuration is found to well improve the sparsity of dictionaries, resulting in an enhanced classification performance. Moreover, the results show that the proposed method is better than conventional SVM classifier for classifying breast masses subject to various margins from normal tissues.

Multiresolution local binary pattern texture analysis combined with variable selection for application to false-positive reduction in computer-aided detection of breast masses on mammograms.

In this paper, a new and novel approach is designed for extracting local binary pattern (LBP) texture features from the computer-identified mass regions, aiming to reduce false-positive (FP) detection in a computerized mass detection framework. The proposed texture feature, the so-called multiresolution LBP feature, is well able to characterize the regional texture patterns of core and margin regions of a mass, as well as to preserve the spatial structure information of the mass. In addition, to maximize a complementary effect on improving classification accuracy, multiresolution texture analysis has been incorporated into the extraction of LBP features. Further, SVM-RFE-based variable selection strategy is applied for selecting an optimal subset of variables of multiresolution LBP texture features to maximize the separation between breast masses and normal tissues. Extensive and comparative experiments have been conducted to evaluate the proposed method on two public benchmark mammogram databases (DBs). Experimental results show that the proposed multiresolution LBP features (extracted from automatically segmented mass boundaries) outperform other state-of-the-art texture features developed for FP reduction. Our results also indicate that combining our multiresolution LBP features with variable selection strategy is an effective solution for reducing FP signals in computer-aided detection (CAD) of mammographic masses.

Combining multiple feature representations and AdaBoost ensemble learning for reducing false-positive detections in computer-aided detection of masses on mammograms.

One of the drawbacks of current Computer-aided Detection (CADe) systems is a high number of false-positive (FP) detections, especially for detecting mass abnormalities. In a typical CADe system, classifier design is one of the key steps for determining FP detection rates. This paper presents the effective classifier ensemble system for tackling FP reduction problem in CADe. To construct ensemble consisting of correct classifiers while disagreeing with each other as much as possible, we develop a new ensemble construction solution that combines data resampling underpinning AdaBoost learning with the use of different feature representations. In addition, to cope with the limitation of weak classifiers in conventional AdaBoost, our method has an effective mechanism for tuning the level of weakness of base classifiers. Further, for combining multiple decision outputs of ensemble members, a weighted sum fusion strategy is used to maximize a complementary effect for correct classification. Comparative experiments have been conducted on benchmark mammogram dataset. Results show that the proposed classifier ensemble outperforms the best single classifier in terms of reducing the FP detections of masses.