Ultrasound Images Of Breast Cancer Research Articles

AMSLS: Adaptive multi-scale level set method based on local entropy for image segmentation

Intensity inhomogeneity often appears in medical images and causes great difficulties in image segmentation. Most active contour models perform poorly when applied to intensity inhomogeneous images because their energy functions use local intensity information in a fixed-size domain, causing the contour to evolve in the wrong direction. To overcome the difficulties caused by intensity inhomogeneity, we propose an adaptive multi-scale Gaussian kernel function based on image local entropy, which can determine the appropriate scale for each local region. Choosing the small scale and large scale for inhomogeneous and homogeneous areas respectively make the contour move toward the target boundary accurately. We also propose three adaptive multi-scale (AMS) models, AMS-region scalable fitting (AMS-RSF) model, AMS-local image fitting (AMS-LIF) model, AMS-local and global intensity fitting (AMS-LGIF) model, to segment medical images with intensity inhomogeneity and noise, including left atrial MR images and breast ultrasound images. The experimental results show that the adaptive multi-scale Gaussian kernel function enables the active contour model to effectively segment intensity inhomogeneous images and has a certain robustness to the initial contour and noise, which achieves good performance on MR left atrial images and ultrasound images of breast cancer. The AMS-LGIF model obtained the highest DICE coefficient of 0.9532, which was better than the 0.9429 obtained by the second-ranked LGIF model to segment left atrial MR images. For segmenting breast ultrasound images, the DICE coefficient is increased by 16% than that of the U-Net++ model.

Revolutionizing breast cancer Ki-67 diagnosis: ultrasound radiomics and fully connected neural networks (FCNN) combination method.

This study aims to assess the diagnostic value of ultrasound habitat sub-region radiomics feature parameters using a fully connected neural networks (FCNN) combination method L2,1-norm in relation to breast cancer Ki-67 status. Ultrasound images from 528 cases of female breast cancer at the Affiliated Hospital of Xiangnan University and 232 cases of female breast cancer at the Affiliated Rehabilitation Hospital of Xiangnan University were selected for this study. We utilized deep learning methods to automatically outline the gross tumor volume and perform habitat clustering. Subsequently, habitat sub-regions were extracted to identify radiomics features and underwent feature engineering using the L1,2-norm. A prediction model for the Ki-67 status of breast cancer patients was then developed using a FCNN. The model's performance was evaluated using accuracy, area under the curve (AUC), specificity (Spe), positive predictive value (PPV), negative predictive value (NPV), Recall, and F1. In addition, calibration curves and clinical decision curves were plotted for the test set to visually assess the predictive accuracy and clinical benefit of the models. Based on the feature engineering using the L1,2-norm, a total of 9 core features were identified. The predictive model, constructed by the FCNN model based on these 9 features, achieved the following scores: ACC 0.856, AUC 0.915, Spe 0.843, PPV 0.920, NPV 0.747, Recall 0.974, and F1 0.890. Furthermore, calibration curves and clinical decision curves of the validation set demonstrated a high level of confidence in the model's performance and its clinical benefit. Habitat clustering of ultrasound images of breast cancer is effectively supported by the combined implementation of the L1,2-norm and FCNN algorithms, allowing for the accurate classification of the Ki-67 status in breast cancer patients.

Few-shot classification of ultrasound breast cancer images using meta-learning algorithms

Medical datasets often have a skewed class distribution and a lack of high-quality annotated images. However, deep learning methods require a large amount of labeled data for classification. In this study, we present a few-shot learning approach for the classification of ultrasound breast cancer images using meta-learning methods. We used prototypical networks and model agnostic meta-learning (MAML) algorithms as meta-learning methods. The breast ultrasound images (BUSI) dataset, which has three classes and is difficult to use in meta-learning, was used for meta-testing in a cross-domain approach along with other datasets for meta-training. Our proposed approach yielded an accuracy range of 0.882–0.889, achieved by implementing the ResNet50 backbone with ProtoNet in a 10-shot setting. These results represent a significant improvement ranging from 6.27 to 7.10% over the baseline accuracy of 0.831. The results showed that ProtoNet outperformed the MAML method for all k-shot settings. In addition, the use of ResNet models as the backbone network for feature extraction was found to be more successful than the use of a four-layer convolutional model. Our proposed method is the first attempt to apply meta-learning for few-shot classification in the BUSI dataset while providing higher accuracy compared to deep learning methods for medical images with small-scale datasets and few classes. The methodology used in this study can be adapted to other datasets with similar problems.

When deep learning is not enough: artificial life as a supplementary tool for segmentation of ultrasound images of breast cancer.

Segmentation of tumors in ultrasound (US) images of the breast is a critical issue in medical imaging. Due to the poor quality of US images and the varying specifications of US machines, segmentation and classification of abnormalities present difficulties even for trained radiologists. The paper aims to introduce a novel AI-based hybrid model for US segmentation that offers high accuracy, requires relatively smaller datasets, and is capable of handling previously unseen data. The software can be used for diagnostics and the US-guided biopsies. A unique and robust hybrid approach that combines deep learning (DL) and multi-agent artificial life (AL) has been introduced. The algorithms are verified on three US datasets. The method outperforms 14 selected state-of-the-art algorithms applied to US images characterized by complex geometry and high level of noise. The paper offers an original classification of the images and tests to analyze the limits of the DL. The model has been trained and verified on 1264 ultrasound images. The images are in the JPEG and PNG formats. The age of the patients ranges from 22 to 73 years. The 14 benchmark algorithms include deformable shapes, edge linking, superpixels, machine learning, and DL methods. The tests use eight-region shape- and contour-based evaluation metrics. The proposed method (DL-AL) produces excellent results in terms of the dice coefficient (region) and the relative Hausdorff distance H3 (contour-based) as follows: the easiest image complexity level, Dice = 0.96 and H3 = 0.26; the medium complexity level, Dice = 0.91 and H3 = 0.82; and the hardest complexity level, Dice = 0.90 and H3 = 0.84. All other metrics follow the same pattern. The DL-AL outperforms the second best (Unet-based) method by 10-20%. The method has been also tested by a series of unconventional tests. The model was trained on low complexity images and applied to the entire set of images. These results are summarized below. (1) Only the low complexity images have been used for training (68% unknown images): Dice = 0.80 and H3 = 2.01. (2) The low and the medium complexity images have been used for training (51% unknown images): Dice = 0.86 and H3 = 1.32. (3) The low, medium, and hard complexity images have been used for training (35% unknown images): Dice = 0.92 and H3 = 0.76. These tests show a significant advantage of DL-AL over 30%. A video demo illustrating the algorithm is at http://tinyurl.com/mr4ah687 .

Breast Tumor Recognition by Semantic Segmentation of Multiclass Ultrasound Images

Objectives: The main purpose of this paper is to suggest a semantic segmentation model to reduce training time in ultrasound breast cancer images. This is achieved by employing a smaller network with fewer trainable parameters, resulting in faster training while maintaining maximum accuracy.
 Methods: This paper proposes a modified U-Net model, which we call the V model, for the subdivision of breast tumors. The proposed V architecture is applied explicitly to ultrasound breast cancer datasets for semantic segmentation. Our proposed model achieves semantic segmentation by employing an encoder and decoder on real and mask image datasets.
 Findings: Therefore, developing a proposed system, namely a V-Net computer-aided diagnosis (CAD) system, is imperative. This CAD system aims to minimize human errors while enhancing accuracy and speed in the premature finding of breast tumors. The proposed model utilizes minimal layers and parameters while maintaining superior results regarding correctness, speed, and computational proficiency.
 Novelty: The proposed V-net model applies to analysing any medical image for detecting disease and finding more accuracy than other U-net models.

Classification of Breast Masses Using Ultrasound Images by Approaching GAN, Transfer Learning and Deep Learning Techniques

Breast cancer is a common cause of death among women worldwide. Ultrasonic imaging is a valuable diagnostic tool in breast cancer detection. However, the accuracy of computer-aided diagnosis systems for breast cancer classification is limited due to the lack of well-annotated datasets. This study proposes a deep learning-based framework for breast mass classification using ultrasound images, which incorporates a novel data augmentation technique, Generative Adversarial Network (GAN), and Transfer Learning (TL). Automating early tumor identification and classification in breast cancer diagnosis can save lives by improving the accuracy of diagnoses and reducing the need for invasive procedures. However, the limited availability of well-annotated datasets for ultrasound images of breast cancer has hampered the development of accurate computer-aided diagnosis systems. The accuracy of breast mass classification using ultrasound images is limited due to the lack of well-annotated datasets. Conventional data augmentation techniques have limitations in applications with strict guidelines, such as medical datasets. Therefore, there is a need to develop a novel data augmentation technique to improve the accuracy of breast mass classification using ultrasound images. The proposed framework can be extended to other medical imaging applications, where the availability of well-annotated datasets is limited. The GAN-based data augmentation technique and TL-based feature extraction can be used to improve the accuracy of classification models in other medical imaging applications. Additionally, the proposed framework can be used to develop accurate computer-aided diagnosis systems for breast cancer detection in clinical settings. The proposed framework incorporates a deep learning-based approach for breast mass classification using ultrasound images. The framework includes a GAN-based data augmentation technique and TL for feature extraction. The dataset used for training and testing the model is the Breast Ultrasound Images (BUSI) dataset, which includes 1311 images with normal and abnormal breast masses. The proposed framework achieved an accuracy of 99.6% for breast mass classification using ultrasound images, which outperformed existing methods. The GAN-based data augmentation technique and TL-based feature extraction improved the accuracy of the classification model. The results suggest that deep learning algorithms can be effectively applied for breast ultrasound categorization. The proposed framework presents a novel approach for breast mass classification using ultrasound images, which incorporates a GAN-based data augmentation technique and TL-based feature extraction. The results demonstrate that the proposed framework outperforms existing methods and achieves high accuracy in breast mass classification using ultrasound images. This framework can be useful for developing accurate computer-aided diagnosis systems for breast cancer detection.

Enhanced Computer-Aided Diagnosis Model on Ultrasound Images through Transfer Learning and Data Augmentation Techniques for an Accurate Breast Tumors Classification

Cancer is a critical global public health problem with meager median survival. It is therefore quite essential to detect this disease at an early stage to improve diagnostic results and consequently avoid serious complications. For this purpose, various researchers have implemented automated methods with the use of different medical imaging modalities. Accordingly, the expansion of deep learning techniques grants opportunities to enhance diagnosis, cure, and prevention. In this study, a diagnostic system for accurate classification of ultrasound breast abnormalities based on the powerful ResNet-50 CNN is proposed with the aim of providing early detection of breast cancer decease. The contribution of this work lies in the novel approach taken to improve the performance of the ResNet50 model in the classification of ultrasound breast cancer images. Transfer learning allows for the model to leverage pre-existing knowledge, while the application of data augmentation techniques enhances the diversity and quality of the training data. Additionally, the optimization of the batch size as a hyperparameter ensures that the model is able to effectively learn from the training data, leading to improved accuracy and efficiency in the classification process. This approach is crucial in the early detection and treatment of breast cancer. Quantitative and qualitative evaluations have been detailed in this study using Breast Ultrasound Dataset BUSI. Our presented work shows interesting results in terms of accuracy, specificity, sensitivity, and AUC which exceed the performance of other compared works. Moreover, the proposed method helps boost the clinical diagnosis of breast cancer. It may integrate a radiologist network, allowing them to constantly follow up on the patient's medical history.

Automation of ultrasound breast cancer images classification using deep neural networks

Automation of ultrasound breast cancer images classification using deep neural networks

Artificial life for segmentation of fusion ultrasound images of breast abnormalities

Segmentation of cancerous tumors in ultrasound (US) images of human organs is one of the critical problems in medical imaging. The US images are characterized by low contrast, irregular shapes, high levels of speckle-noise and acoustic shadows, making it difficult to segment the tumor. Yet, US imaging is considered one of the most inexpensive and safe imaging tests available to detect cancer in its early stages. However, an automatic segmentation method applicable to all types of US imagery does not exist.This paper proposes a novel segmentation method that combines image fusion, artificial life (AL) and a genetic algorithm (GA). The new algorithm has been applied to US images of breast cancer. The method is based on tracing agents (TA), which are artificial organisms with memory and the ability to communicate. They live inside a fusion image generated from the US and the elastography (EL) images.The TA can recognize the patterns of strong edges and boundary gaps allowing to outline the tumor. The new model has been tested against six types of segmentation models, i.e., machine learning, active contours, level set models, superpixel models, edge linking models and selected hybrid methods. The experiments include 16 state-of-the-art methods, which outperform 69 recent and classical segmentation routines. The tests were run on 395 breast cancer images from http://onlinemedicalimages.com and https://www.ultrasoundcases.info/.TA training employs a GA. The model has been verified on “hard” cases (complex shapes, boundary leakage, and noisy edge maps). The proposed algorithm produces more accurate results than the reference methods on high complexity images. A video demo of the algorithm is at http://shorturl.at/htBW9.

Recognition of Ultrasound Images of Breast Cancer Based on Discrete Sinc Transform and Deep Learning

Recognition of Ultrasound Images of Breast Cancer Based on Discrete Sinc Transform and Deep Learning

Restoration and enhancement of breast ultrasound images using extended complex diffusion based unsharp masking.

Ultrasound is a well-known imaging modality for the interpretation of breast cancer. It is playing very important role for breast cancer detection that are missed by mammograms. The image acquisition is usually affected by the presence of noise, artifacts, and distortion. To overcome such type of issues, there is a need of image restoration and enhancement to improve the quality of image. This paper proposes a single framework for denoising and enhancement of ultrasound images, where a smoothing filter is replaced with an extended complex diffusion-based filter in an unsharp masking technique. The performance evaluation of the proposed method is tested on real ultrasound breast cancer images database and synthetic ultrasound image. The performance evaluation comprises qualitative and quantitative evaluation along with comparative analysis of pre-existing and proposed method. The quantitative evaluation metrics are mean squared error, peak-signal-to-noise ratio, correlation parameter, normalized absolute error, universal quality index, similarity structure index, edge preservation index, a measure of enhancement, a measure of enhancement by entropy, and second derivative like measurement. The result specifies that the proposed method is better suited approach for the removal of speckle noise which follows Rayleigh distribution, restoration of information, enhancement of abnormalities, and proper edge preservation.

Abstract PS3-29: Pixel-level tissue classification from ultrasound images of breast cancer and direct comparison to matched histological measurements

Abstract Introduction: The use of noninvasive ultrasound (US) for quantitative tissue characterization has been an exciting research prospect for several decades now. Herein the challenge is to find hidden patterns in the US data to reveal more information about tissue function and pathology that cannot be seen in the more conventional US images. A new pixel-level analysis technique has been developed by our group for tissue classification. Termed H-scan US (H stands for hue or Hermite), this imaging approach links a special class of nth-ordered Gaussian-weighted Hermite functions (GHn) to the physics of US scattering and reflection from different tissue structures. The sensitivity of in vivo H-scan US imaging to subtle changes in scatterer size is a question that has not been fully answered. To that end, the purpose of this study was to compare local H-scan US image intensity to direct (co-registered) histological measures made at the cellular level. Methodology Female nude athymic mice were implanted into mammary fatty pad with breast cancer cells (N = 8). Implanted tumors were allowed to grow for about four weeks before study enrollment. Image data was acquired using a programmable US scanner (Vantage 256, Verasonics Inc) equipped with a 256-element L22-8v CMUT linear array transducer (Kolo Medical). Plane wave imaging with 5 angles was performed at a center frequency of 15 MHz. To generate the H-scan US image, three parallel convolution filters (GH2, GH6, and GH10) were applied to the radiofrequency (RF) data sequences to measure the relative strength of the received signals. After envelope detection, the relative strength of the filter outputs is color coded whereby the lower frequency (GH2 = 9 MHz) backscattered US signal components are assigned to the R channel, moderate frequency (GH6 = 15 MHz) signals are assigned to the G channel, and the higher frequency (GH10 = 21 MHz) signals to the B channel. After performing H-scan US, the imaging cross-section was marked, and the tumors were excised and sliced along the same plane for histologic processing. After nuclear staining, tissue sections were scanned and digitized using confocal microscopy and fluorescent filters for DAPI. Automated segmentation of each cancer cell nucleus in the histologic sections was performed using an active contour technique. US images were interpolated to the same number of pixels as the histology image and spatial alignment between the two was performed. Lastly, nucleus size and density from histologic sections was compared to local H-scan US image features. Results: Custom software was developed to compare local US image features at the pixel level to that of co-registered histologic images and measurements of nuclear size and density. In an attempt to properly match the histological and the H-scan US images, three different sized kernels of 0.16 ✗ 0.16 mm, 0.32 ✗ 0.32 mm or 0.5 ✗ 0.5 mm were used to partition both the H-scan US and histology images into 1888, 448 or 62 distinct region-of-interests (ROIs), respectively. Mean nuclear size and density measurements from the histology ROIs was compared to the mean H-scan US image intensity from the many spatially matched ROI locations. A statistically significant linear relationship was found between local H-scan US image intensity and nuclear size (R2 > 0.4, < 0.001) and density (R2 > 0.6, p < 0.001). Conclusions: Preliminary results from use of an animal model of breast cancer reveals that in vivo H-scan US images positively correlated with physical measures of nucleus size and density as quantified from co-registered histologic images. Citation Format: Mawia Khairalseed, Shreya Reddy, Jane Song, Girdhari Rijal, Kenneeth Hoyt. Pixel-level tissue classification from ultrasound images of breast cancer and direct comparison to matched histological measurements [abstract]. In: Proceedings of the 2020 San Antonio Breast Cancer Virtual Symposium; 2020 Dec 8-11; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2021;81(4 Suppl):Abstract nr PS3-29.

Multi-Level Fusion in Ultrasound for Cancer Detection based on Uniform LBP Features

Collective improvement in the acceptable or desirable accuracy level of breast cancer image-related pattern recognition using various schemes remains challenging. Despite the combination of multiple schemes to achieve superior ultrasound image pattern recognition by reducing the speckle noise, an enhanced technique is not achieved. The purpose of this study is to introduce a features-based fusion scheme based on enhancement uniform-Local Binary Pattern (LBP) and filtered noise reduction. To surmount the above limitations and achieve the aim of the study, a new descriptor that enhances the LBP features based on the new threshold has been proposed. This paper proposes a multi-level fusion scheme for the auto-classification of the static ultrasound images of breast cancer, which was attained in two stages. First, several images were generated from a single image using the pre-processing method. The median and Wiener filters were utilized to lessen the speckle noise and enhance the ultrasound image texture. This strategy allowed the extraction of a powerful feature by reducing the overlap between the benign and malignant image classes. Second, the fusion mechanism allowed the production of diverse features from different filtered images. The feasibility of using the LBP-based texture feature to categorize the ultrasound images was demonstrated. The effectiveness of the proposed scheme is tested on 250 ultrasound images comprising 100 and 150 benign and malignant images, respectively. The proposed method achieved very high accuracy (98%), sensitivity (98%), and specificity (99%). As a result, the fusion process that can help achieve a powerful decision based on different features produced from different filtered images improved the results of the new descriptor of LBP features in terms of accuracy, sensitivity, and specificity.

An Efficient System for Early Diagnosis of B reast Cancer using Support Vector Machine

There are many lives lost every year due to cancer and among them; among the women breast cancer causes the most deaths. For the better prediction of breast cancer risks, numerous studies have been undertaken incorporating data mining techniques. 1.1 million Cases of breast cancer were reported in 2004. It has been seen over the years that, that the numbers increase with the increasing industrialization and urbanization. It was earlier observed that mostly affected countries with breast cancer were high income countries such as America but now a days it is also very serious issue in middle and low income countries like Africa, Latin America and Asia. The main objective of this paper is to create a model which can more efficiently and accurately categorize a cancer as malignant or benevolent based on interpretation of the numerical values of attributes of ultrasound images of breast cancer. In this paper various data mining algorithm used like SVM(Support Vector Machine) for prediction and compared it with various other algorithms such as CART, Logistic Regression, KNN for the best training and test accuracy. SVM algorithm gives the most accurate results among the rest algorithm.

Deep Learning Approaches for Data Augmentation and Classification of Breast Masses using Ultrasound Images

Breast classification and detection using ultrasound imaging is considered a significant step in computer-aided diagno-sis systems. Over the previous decades, researchers have proved the opportunities to automate the initial tumor classification and detection. The shortage of popular datasets of ultrasound images of breast cancer prevents researchers from obtaining a good performance of the classification algorithms. Traditional augmentation approaches are firmly limited, especially in tasks where the images follow strict standards, as in the case of medical datasets. Therefore besides the traditional augmentation, we use a new methodology for data augmentation using Generative Adversarial Network (GAN). We achieved higher accuracies by integrating traditional with GAN-based augmentation. This paper uses two breast ultrasound image datasets obtained from two various ultrasound systems. The first dataset is our dataset which was collected from Baheya Hospital for Early Detection and Treatment of Women’s Cancer, Cairo (Egypt), we name it (BUSI) referring to Breast Ultrasound Images (BUSI) dataset. It contains 780 images (133 normal, 437 benign and 210 malignant). While the Dataset (B) is obtained from related work and it has 163 images (110 benign and 53 malignant). To overcome the shortage of public datasets in this field, BUSI dataset will be publicly available for researchers. Moreover, in this paper, deep learning approaches are proposed to be used for breast ultrasound classification. We examine two different methods: a Convolutional Neural Network (CNN) approach and a Transfer Learning (TL) approach and we compare their performance with and without augmentation. The results confirm an overall enhancement using augmentation methods with deep learning classification methods (especially transfer learning) when evaluated on the two datasets.

A novel classifier model for mass classification using BI-RADS category in ultrasound images based on Type-2 fuzzy inference system

Ultrasound imaging is an imaging technique for early detection of breast cancer. Breast Imaging Reporting and Data System (BI-RADS) lexicon, developed by The American College of Radiology, provides a standard for expert doctors to interpret the ultrasound images of breast cancer. This standard describes the features to classify the tumour as benign or malignant and it also categorizes the biopsy requirement as a percentage. Biopsy is an invasive method that doctors use for diagnosis of breast cancer. Computer-aided detection (CAD)/diagnosis systems that are designed to include the feature standards used in benign/malignant classification help the doctors in diagnosis but they do not provide enough information about the BI-RADS category of the mass. These systems classify the benign tumours with 90% biopsy possibility (BI-RADS-4) and with 2% biopsy possibility (BI-RADS-2) in the same category. There are some studies in the literature that make category classification via commonly used classifier methods but their success rates are low. In this study, a two-layer, high-success-rate classifier model based on Type-2 fuzzy inference is developed, which classifies the tumour as benign or malignant with its BI-RADS category by incorporating the opinions of the expert doctors. A 99.34% success rate in benign/malignant classification and a 92% success rate in category classification (BI-RADS 2, 3, 4, 5) were obtained in the accuracy tests. These results indicate that the CAD system is valuable as a means of providing a second diagnostic opinion when radiologists carry out mass diagnosis.

Automatic initialization of active contours in ultrasound images of breast cancer

Gradient vector flow (GVF) snakes are an efficient method for segmentation of ultrasound images of breast cancer. However, the method produces inaccurate results if the seeds are initialized improperly (far from the true boundaries and close to the false boundaries). Therefore, we propose a novel initialization method designed for GVF-type snakes based on walking particles. At the first step, the algorithm locates the seeds at converging and diverging configurations of the vector field. At the second step, the seeds "explode," generating a set of random walking particles designed to differentiate between the seeds located inside and outside the object. The method has been tested against five state-of-the-art initialization methods on sixty ultrasound images from a database collected by Thammasat University Hospital of Thailand (http://onlinemedicalimages.com). The ground truth was hand-drawn by leading radiologists of the hospital. The competing methods were: trial snake method (TS), centers of divergence method (CoD), force field segmentation (FFS), Poisson Inverse Gradient Vector Flow (PIG), and quasi-automated initialization (QAI). The numerical tests demonstrated that CoD and FFS failed on the selected test images, whereas the average accuracy of PIG and QAI was 73 and 87%, respectively, versus 97% achieved by the proposed method. Finally, TS has shown a comparable accuracy of about 93%; however, the method is about ten times slower than the proposed exploding seeds. A video demonstration of the algorithm is at http://onlinemedicalimages.com/index.php/en/presentations.

Reproducibility of quantitative high-throughput BI-RADS features extracted from ultrasound images of breast cancer.

Digital Breast Imaging Reporting and Data System (BI-RADS) features extracted from ultrasound images are essential in computer-aided diagnosis, prediction, and prognosis of breast cancer. This study focuses on the reproducibility of quantitative high-throughput BI-RADS features in the presence of variations due to different segmentation results, various ultrasound machine models, and multiple ultrasound machine settings. Dataset 1 consists of 399 patients with invasive breast cancer and is used as the training set to measure the reproducibility of features, while dataset 2 consists of 138 other patients and is a validation set used to evaluate the diagnosis performances of the final reproducible features. Four hundred and sixty high-throughput BI-RADS features are designed and quantized according to BI-RADS lexicon. Concordance Correlation Coefficient (CCC) and Deviation (Dev) are used to assess the effect of the segmentation methods and Between-class Distance (BD) is used to study the influences of the machine models. In addition, the features jointly shared by two methodologies are further investigated on their effects with multiple machine settings. Subsequently, the absolute value of Pearson Correlation Coefficient (Rabs ) is applied for redundancy elimination. Finally, the features that are reproducible and not redundant are preserved as the stable feature set. A 10-fold Support Vector Machine (SVM) classifier is employed to verify the diagnostic ability. One hundred and fifty-three features were found to have high reproducibility (CCC > 0.9 & Dev < 0.1) within the manual and automatic segmentation. Three hundred and thirty-nine features were stable (BD < 0.2) at different machine models. Two feature sets shared the same 102 features, in which nine features were highly sensitive to the machine settings. Forty-six features were finally preserved after redundancy elimination. For the validation in dataset 2, the area under curve (AUC) of the 10-fold SVM classifier was 0.915. Three factors, segmentation results, machine models, and machine settings may affect the reproducibility of high-throughput BI-RADS features to various degrees. Our 46 reproducible features were robust to these factors and were capable of distinguishing benign and malignant breast tumors.

MO-DE-207B-06: Computer-Aided Diagnosis of Breast Ultrasound Images Using Transfer Learning From Deep Convolutional Neural Networks

Purpose: To assess the performance of using transferred features from pre-trained deep convolutional networks (CNNs) in the task of classifying cancer in breast ultrasound images, and to compare this method of transfer learning with previous methods involving human-designed features. Methods: A breast ultrasound dataset consisting of 1125 cases and 2393 regions of interest (ROIs) was used. Each ROI was labeled as cystic, benign, or malignant. Features were extracted from each ROI using pre-trained CNNs and used to train support vector machine (SVM) classifiers in the tasks of distinguishing non-malignant (benign+cystic) vs malignant lesions and benign vs malignant lesions. For a baseline comparison, classifiers were also trained on prior analytically-extracted tumor features. Five-fold cross-validation (by case) was conducted with the area under the receiver operating characteristic curve (AUC) as the performance metric. Results: Classifiers trained on CNN-extracted features were comparable to classifiers trained on human-designed features. In the non-malignant vs malignant task, both the SVM trained on CNN-extracted features and the SVM trained on human-designed features obtained an AUC of 0.90. In the task of determining benign vs malignant, the SVM trained on CNN-extracted features obtained an AUC of 0.88, compared to the AUC of 0.85 obtained by the SVM trained on human-designed features. Conclusion: We obtained strong results using transfer learning to characterize ultrasound breast cancer images. This method allows us to directly classify a small dataset of lesions in a computationally inexpensive fashion without any manual input. Modern deep learning methods in computer vision are contingent on large datasets and vast computational resources, which are often inaccessible for clinical applications. Consequently, we believe transfer learning methods will be important for computer-aided diagnosis schemes in order to utilize advancements in deep learning and computer vision without the associated costs. This work was partially funded by NIH grant U01 CA195564 and the University of Chicago Metcalf program. M.L.G. is a stockholder in R2/Hologic, co-founder and equity holder in Quantitative Insights, and receives royalties from Hologic, GE Medical Systems, MEDIAN Technologies, Riverain Medical, Mitsubishi, and Toshiba. K.D. received royalties from Hologic.

Phase portrait analysis for automatic initialization of multiple snakes for segmentation of the ultrasound images of breast cancer

Segmentation of ultrasound (US) images of breast cancer is one of the most challenging problems of modern medical image processing. A number of popular codes for US segmentation are based on the active contours (snakes) and on a variety of modifications of gradient vector flow. The snakes have been used to locate objects in various applications of medical images. However, the main difficulty in applying the method is initialization. Therefore, we suggest a new method for automatic initialization of active contours based on phase portrait analysis (PPA) of the underlying vector field and a sequential initialization of trial multiple snakes. The PPA makes it possible to exclude the noise and artifacts and properly initialize the multiple snakes. In turn, the trial snakes allow us to differentiate between the seeds initialized inside and outside the desired object. While preceding methods require the manual selection of at least one seed point inside the object or rely on the particular distribution of the gray levels, the proposed method is fully automatic and robust to the noise, as can be seen from the tests with synthetic and real images.