Feature extraction and classification of breast cancer on dynamic magnetic resonance imaging using artificial neural network

A multicentric study of radiomics and artificial intelligence analysis on contrast-enhanced mammography to identify different histotypes of breast cancer

OBJECTIVE

To evaluate the performance of radiomic analysis on contrast-enhanced mammography images to identify different histotypes of breast cancer mainly in order to predict grading, to identify hormone receptors, to discriminate human epidermal growth factor receptor 2 (HER2) and to identify luminal histotype of the breast cancer.

METHODS

From four Italian centers were recruited 180 malignant lesions and 68 benign lesions. However, only the malignant lesions were considered for the analysis. All patients underwent contrast-enhanced mammography in cranium caudal (CC) and medium lateral oblique (MLO) view. Considering histological findings as the ground truth, four outcomes were considered: (1) G1 + G2 vs. G3; (2) HER2 + vs. HER2 - ; (3) HR + vs. HR - ; and (4) non-luminal vs. luminal A or HR + /HER2-  and luminal B or HR + /HER2 + . For multivariate analysis feature selection, balancing techniques and patter recognition approaches were considered.

RESULTS

The univariate findings showed that the diagnostic performance is low for each outcome, while the results of the multivariate analysis showed that better performances can be obtained. In the HER2 + detection, the best performance (73% of accuracy and AUC = 0.77) was obtained using a linear regression model (LRM) with 12 features extracted by MLO view. In the HR + detection, the best performance (77% of accuracy and AUC = 0.80) was obtained using a LRM with 14 features extracted by MLO view. In grading classification, the best performance was obtained by a decision tree trained with three predictors extracted by MLO view reaching an accuracy of 82% on validation set. In the luminal versus non-luminal histotype classification, the best performance was obtained by a bagged tree trained with 15 predictors extracted by CC view reaching an accuracy of 94% on validation set.

CONCLUSIONS

The results suggest that radiomics analysis can be effectively applied to design a tool to support physician decision making in breast cancer classification. In particular, the classification of luminal versus non-luminal histotypes can be performed with high accuracy.

Radiomics and artificial intelligence analysis by T2-weighted imaging and dynamic contrast-enhanced magnetic resonance imaging to predict Breast Cancer Histological Outcome

OBJECTIVE

The objective of the study was to evaluate the accuracy of radiomics features obtained by MR images to predict Breast Cancer Histological Outcome.

METHODS

A total of 217 patients with malignant lesions were analysed underwent MRI examinations. Considering histological findings as the ground truth, four different types of findings were used in both univariate and multivariate analyses: (1) G1 + G2 vs G3 classification; (2) presence of human epidermal growth factor receptor 2 (HER2 + vs HER2 -); (3) presence of the hormone receptor (HR + vs HR -); and (4) presence of luminal subtypes of breast cancer.

RESULTS

The best accuracy for discriminating HER2 + versus HER2 - breast cancers was obtained considering nine predictors by early phase T1-weighted subtraction images and a decision tree (accuracy of 88% on validation set). The best accuracy for discriminating HR + versus HR - breast cancers was obtained considering nine predictors by T2-weighted subtraction images and a decision tree (accuracy of 90% on validation set). The best accuracy for discriminating G1 + G2 versus G3 breast cancers was obtained considering 16 predictors by early phase T1-weighted subtraction images in a linear regression model with an accuracy of 75%. The best accuracy for discriminating luminal versus non-luminal breast cancers was obtained considering 27 predictors by early phase T1-weighted subtraction images and a decision tree (accuracy of 94% on validation set).

CONCLUSIONS

The combination of radiomics analysis and artificial intelligence techniques could be used to support physician decision-making in prediction of Breast Cancer Histological Outcome.

Prediction of Breast Cancer Histological Outcome by Radiomics and Artificial Intelligence Analysis in Contrast-Enhanced Mammography

Simple Summary

The assessment of breast lesions through mammographic images is currently challenging, especially in dense breasts. Contrast-enhanced mammography has been shown to overcome the limitations of standard mammography but it greatly depends on the interpretative skills of the physician. The aim of this study was to evaluate the potentialities of statistical and artificial intelligence algorithms as a tool for helping the radiologists in the interpretation of images. The most remarkable results were achieved in discriminating benign from malignant lesions and in the identification of the presence of the hormone receptor. A tool to support the physician’s decision-making process may be designed starting from simple logistic regression and tree-based algorithms. This type of tool may help the radiologist in assessing the investigated breast and in choosing the appropriate follow-up without resorting to histology.

Abstract

Purpose: To evaluate radiomics features in order to: differentiate malignant versus benign lesions; predict low versus moderate and high grading; identify positive or negative hormone receptors; and discriminate positive versus negative human epidermal growth factor receptor 2 related to breast cancer. Methods: A total of 182 patients with known breast lesions and that underwent Contrast-Enhanced Mammography were enrolled in this retrospective study. The reference standard was pathology (118 malignant lesions and 64 benign lesions). A total of 837 textural metrics were extracted by manually segmenting the region of interest from both craniocaudally (CC) and mediolateral oblique (MLO) views. Non-parametric Wilcoxon–Mann–Whitney test, receiver operating characteristic, logistic regression and tree-based machine learning algorithms were used. The Adaptive Synthetic Sampling balancing approach was used and a feature selection process was implemented. Results: In univariate analysis, the classification of malignant versus benign lesions achieved the best performance when considering the original_gldm_DependenceNonUniformity feature extracted on CC view (accuracy of 88.98%). An accuracy of 83.65% was reached in the classification of grading, whereas a slightly lower value of accuracy (81.65%) was found in the classification of the presence of the hormone receptor; the features extracted were the original_glrlm_RunEntropy and the original_gldm_DependenceNonUniformity, respectively. The results of multivariate analysis achieved the best performances when using two or more features as predictors for classifying malignant versus benign lesions from CC view images (max test accuracy of 95.83% with a non-regularized logistic regression). Considering the features extracted from MLO view images, the best test accuracy (91.67%) was obtained when predicting the grading using a classification-tree algorithm. Combinations of only two features, extracted from both CC and MLO views, always showed test accuracy values greater than or equal to 90.00%, with the only exception being the prediction of the human epidermal growth factor receptor 2, where the best performance (test accuracy of 89.29%) was obtained with the random forest algorithm. Conclusions: The results confirm that the identification of malignant breast lesions and the differentiation of histological outcomes and some molecular subtypes of tumors (mainly positive hormone receptor tumors) can be obtained with satisfactory accuracy through both univariate and multivariate analysis of textural features extracted from Contrast-Enhanced Mammography images.

Radiomic and Artificial Intelligence Analysis with Textural Metrics Extracted by Contrast-Enhanced Mammography and Dynamic Contrast Magnetic Resonance Imaging to Detect Breast Malignant Lesions

Purpose:The purpose of this study was to discriminate between benign and malignant breast lesions through several classifiers using, as predictors, radiomic metrics extracted from CEM and DCE-MRI images. In order to optimize the analysis, balancing and feature selection procedures were performed. Methods: Fifty-four patients with 79 histo-pathologically proven breast lesions (48 malignant lesions and 31 benign lesions) underwent both CEM and DCE-MRI. The lesions were retrospectively analyzed with radiomic and artificial intelligence approaches. Forty-eight textural metrics were extracted, and univariate and multivariate analyses were performed: non-parametric statistical test, receiver operating characteristic (ROC) and machine learning classifiers. Results: Considering the single metrics extracted from CEM, the best predictors were KURTOSIS (area under ROC curve (AUC) = 0.71) and SKEWNESS (AUC = 0.71) calculated on late MLO view. Considering the features calculated from DCE-MRI, the best predictors were RANGE (AUC = 0.72), ENERGY (AUC = 0.72), ENTROPY (AUC = 0.70) and GLN (gray-level nonuniformity) of the gray-level run-length matrix (AUC = 0.72). Considering the analysis with classifiers and an unbalanced dataset, no significant results were obtained. After the balancing and feature selection procedures, higher values of accuracy, specificity and AUC were reached. The best performance was obtained considering 18 robust features among all metrics derived from CEM and DCE-MRI, using a linear discriminant analysis (accuracy of 0.84 and AUC = 0.88). Conclusions: Classifiers, adjusted with adaptive synthetic sampling and feature selection, allowed for increased diagnostic performance of CEM and DCE-MRI in the differentiation between benign and malignant lesions.

Radiomics and Artificial Intelligence Analysis with Textural Metrics Extracted by Contrast-Enhanced Mammography in the Breast Lesions Classification

The aim of the study was to estimate the diagnostic accuracy of textural features extracted by dual-energy contrast-enhanced mammography (CEM) images, by carrying out univariate and multivariate statistical analyses including artificial intelligence approaches. In total, 80 patients with known breast lesion were enrolled in this prospective study according to regulations issued by the local Institutional Review Board. All patients underwent dual-energy CEM examination in both craniocaudally (CC) and double acquisition of mediolateral oblique (MLO) projections (early and late). The reference standard was pathology from a surgical specimen for malignant lesions and pathology from a surgical specimen or fine needle aspiration cytology, core or Tru-Cut needle biopsy, and vacuum assisted breast biopsy for benign lesions. In total, 104 samples of 80 patients were analyzed. Furthermore, 48 textural parameters were extracted by manually segmenting regions of interest. Univariate and multivariate approaches were performed: non-parametric Wilcoxon–Mann–Whitney test; receiver operating characteristic (ROC), linear classifier (LDA), decision tree (DT), k-nearest neighbors (KNN), artificial neural network (NNET), and support vector machine (SVM) were utilized. A balancing approach and feature selection methods were used. The univariate analysis showed low accuracy and area under the curve (AUC) for all considered features. Instead, in the multivariate textural analysis, the best performance considering the CC view (accuracy (ACC) = 0.75; AUC = 0.82) was reached with a DT trained with leave-one-out cross-variation (LOOCV) and balanced data (with adaptive synthetic (ADASYN) function) and a subset of three robust textural features (MAD, VARIANCE, and LRLGE). The best performance (ACC = 0.77; AUC = 0.83) considering the early-MLO view was reached with a NNET trained with LOOCV and balanced data (with ADASYN function) and a subset of ten robust features (MEAN, MAD, RANGE, IQR, VARIANCE, CORRELATION, RLV, COARSNESS, BUSYNESS, and STRENGTH). The best performance (ACC = 0.73; AUC = 0.82) considering the late-MLO view was reached with a NNET trained with LOOCV and balanced data (with ADASYN function) and a subset of eleven robust features (MODE, MEDIAN, RANGE, RLN, LRLGE, RLV, LZLGE, GLV_GLSZM, ZSV, COARSNESS, and BUSYNESS). Multivariate analyses using pattern recognition approaches, considering 144 textural features extracted from all three mammographic projections (CC, early MLO, and late MLO), optimized by adaptive synthetic sampling and feature selection operations obtained the best results (ACC = 0.87; AUC = 0.90) and showed the best performance in the discrimination of benign and malignant lesions.

Decision support system for breast lesions via dynamic contrast enhanced magnetic resonance imaging

The presented study aims to design a computer-aided detection and diagnosis system for breast dynamic contrast enhanced magnetic resonance imaging. In the proposed system, the segmentation task is performed in two stages. The first stage is called breast region segmentation in which adaptive noise filtering, local adaptive thresholding, connected component analysis, integral of horizontal projection, and breast region of interest detection algorithms are applied to the breast images consecutively. The second stage of segmentation is breast lesion detection that consists of 32-class Otsu thresholding and Markov random field techniques. Histogram, gray level co-occurrence matrix and neighboring gray tone difference matrix based feature extraction, Fisher score based feature selection and, tenfold and leave-one-out cross-validation steps are carried out after segmentation to increase the reliability of the designed system while decreasing the computational time. Finally, support vector machines, k- nearest neighbor, and artificial neural network classifiers are performed to separate the breast lesions as benign and malignant. The average accuracy, sensitivity, specificity, and positive predictive values of each classifier are calculated and the best results are compared with the existing similar studies. According to the achieved results, the proposed decision support system for breast lesion segmentation distinguishes the breast lesions with 86%, 100%, 67%, and 85% accuracy, sensitivity, specificity, and positive predictive values, respectively. These results show that the proposed system can be used to support the radiologists during a breast cancer diagnosis.

Empirical assessment of bias in machine learning diagnostic test accuracy studies

OBJECTIVE

Machine learning (ML) diagnostic tools have significant potential to improve health care. However, methodological pitfalls may affect diagnostic test accuracy studies used to appraise such tools. We aimed to evaluate the prevalence and reporting of design characteristics within the literature. Further, we sought to empirically assess whether design features may be associated with different estimates of diagnostic accuracy.

MATERIALS AND METHODS

We systematically retrieved 2 × 2 tables (n = 281) describing the performance of ML diagnostic tools, derived from 114 publications in 38 meta-analyses, from PubMed. Data extracted included test performance, sample sizes, and design features. A mixed-effects metaregression was run to quantify the association between design features and diagnostic accuracy.

RESULTS

Participant ethnicity and blinding in test interpretation was unreported in 90% and 60% of studies, respectively. Reporting was occasionally lacking for rudimentary characteristics such as study design (28% unreported). Internal validation without appropriate safeguards was used in 44% of studies. Several design features were associated with larger estimates of accuracy, including having unreported (relative diagnostic odds ratio [RDOR], 2.11; 95% confidence interval [CI], 1.43-3.1) or case-control study designs (RDOR, 1.27; 95% CI, 0.97-1.66), and recruiting participants for the index test (RDOR, 1.67; 95% CI, 1.08-2.59).

DISCUSSION

Significant underreporting of experimental details was present. Study design features may affect estimates of diagnostic performance in the ML diagnostic test accuracy literature.

CONCLUSIONS

The present study identifies pitfalls that threaten the validity, generalizability, and clinical value of ML diagnostic tools and provides recommendations for improvement.

Pattern Recognition Approaches for Breast Cancer DCE-MRI Classification: A Systematic Review

We performed a systematic review of several pattern analysis approaches for classifying breast lesions using dynamic, morphological, and textural features in dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI). Several machine learning approaches, namely artificial neural networks (ANN), support vector machines (SVM), linear discriminant analysis (LDA), tree-based classifiers (TC), and Bayesian classifiers (BC), and features used for classification are described. The findings of a systematic review of 26 studies are presented. The sensitivity and specificity are respectively 91 and 83 % for ANN, 85 and 82 % for SVM, 96 and 85 % for LDA, 92 and 87 % for TC, and 82 and 85 % for BC. The sensitivity and specificity are respectively 82 and 74 % for dynamic features, 93 and 60 % for morphological features, 88 and 81 % for textural features, 95 and 86 % for a combination of dynamic and morphological features, and 88 and 84 % for a combination of dynamic, morphological, and other features. LDA and TC have the best performance. A combination of dynamic and morphological features gives the best performance.

Diagnostic assessment by dynamic contrast-enhanced and diffusion-weighted magnetic resonance in differentiation of breast lesions under different imaging protocols

BACKGROUND

The apparent diffusion coefficient (ADC) is a highly diagnostic factor in discriminating malignant and benign breast masses in diffusion-weighted magnetic resonance imaging (DW-MRI). The combination of ADC and other pictorial characteristics has improved lesion type identification accuracy. The objective of this study was to reassess the findings on an independent patient group by changing the magnetic field from 1.5-Tesla to 3.0-Tesla.

METHODS

This retrospective study consisted of a training group of 234 female patients, including 85 benign and 149 malignant lesions, imaged using 1.5-Tesla MRI, and a test group of 95 female patients, including 19 benign and 85 malignant lesions, imaged using 3.0-Tesla MRI. The lesion of interest was segmented from the raw image and four sets of measurements describing the morphology, kinetics, DW-MRI, and texture of the pictorial properties of each lesion were obtained. Each lesion was characterized by 28 features in total. Three classical machine-learning algorithms were used to build prediction models on the training group, which evaluated the prognostic performance of the multi-sided features in three scenarios. To reduce information redundancy, five highly diagnostic factors were selected to obtain a compact yet informative characterization of the lesion status.

RESULTS

Three classification models were built on the training of 1.5-Tesla patients and were tested on the independent 3.0-Tesla test group. The following results were found. i) Characterization of breast masses in a multi-sided way dramatically increased prediction performance. The usage of all features gave a higher performance in both sensitivity and specificity than any individual feature groups or their combinations. ii) ADC was a highly effective factor in improving the sensitivity in discriminating malignant from benign masses. iii) Five features, namely ADC, Sum Average, Entropy, Elongation, and Sum Variance, were selected to achieve the highest performance in diagnosis of the 3.0-Tesla patient group.

CONCLUSIONS

The combination of ADC and other multi-sided characteristics can increase the capability of discriminating malignant and benign breast lesions, even under different imaging protocols. The selected compact feature subsets achieved a high diagnostic performance and thus are promising in clinical applications for discriminating lesion type and for personalized treatment planning.

Accuracy of magnetic resonance in suspicious breast lesions: a systematic quantitative review and meta-analysis

Dynamic contrast-enhanced breast magnetic resonance (MR) is a promising emerging technique for evaluating breast lesions. A quantitative systematic review was performed to estimate the accuracy of breast MR in the diagnosis of high-risk breast lesions and breast cancer. A comprehensive search of the Cochrane Library, MEDLINE, CANCERLIT, LILACS, and EMBASE databases was performed from January 1985 to August 2010. The medical subjects heading (MeSH) and text words for the terms "breast neoplasm", "breast lesions", "breast cancer" and "magnetic resonance" were combined with the MeSH term diagnosis ("sensitivity and specificity"). Studies that compared breast MR with paraffin-embedded sections parameters for the diagnosis of breast lesions (benign, high-risk borderline, and breast cancer) were included. Sixty-nine studies were analyzed, which included 9,298 women with 9,884 breast lesions. Interrater overall agreement between breast MR and paraffin section diagnosis was 79% (κ = 0.55), indicating moderate agreement. Pooled sensitivity and specificity were 90% [95% CI 88-92%] and 75% [95% CI 70-79%], respectively. The pooled likelihood positive ratio was 3.64 (95% CI 3.0-4.2) and the negative ratio was 0.12 (95% CI 0.09-0.15). For breast cancer or high-risk lesions versus benign lesions, the AUC was 0.91 for breast MR and the point Q* was 0.84. In summary, breast MR is a useful pre-operative test for predicting the diagnosis of breast lesions.

Model-Free Visualization of Suspicious Lesions in Breast MRI Based on Supervised and Unsupervised Learning

Dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) has become an important tool in breast cancer diagnosis, but evaluation of multitemporal 3D image data holds new challenges for human observers. To aid the image analysis process, we apply supervised and unsupervised pattern recognition techniques for computing enhanced visualizations of suspicious lesions in breast MRI data. These techniques represent an important component of future sophisticated computer-aided diagnosis (CAD) systems and support the visual exploration of spatial and temporal features of DCE-MRI data stemming from patients with confirmed lesion diagnosis. By taking into account the heterogeneity of cancerous tissue, these techniques reveal signals with malignant, benign and normal kinetics. They also provide a regional subclassification of pathological breast tissue, which is the basis for pseudo-color presentations of the image data. Intelligent medical systems are expected to have substantial implications in healthcare politics by contributing to the diagnosis of indeterminate breast lesions by non-invasive imaging.

Meta-analysis of MR imaging in the diagnosis of breast lesions

PURPOSE

To determine, in a meta-analysis, the diagnostic performance of contrast material-enhanced magnetic resonance (MR) imaging in patients with breast lesions.

MATERIALS AND METHODS

Studies to assess the diagnostic performance of MR imaging in patients suspected of having breast cancer who underwent MR imaging and biopsy from January 1985 through March 2005 were reviewed for inclusion. A summary receiver operating characteristic curve was constructed, and pooled weighted estimates of sensitivity and specificity were calculated by using the recently developed bivariate approach for diagnostic meta-analysis.

RESULTS

Of 251 eligible studies, 44 were included in the meta-analysis (sample size range, 14-821; cancer prevalence, 23%-84%). Pooled weighted estimates of sensitivity and specificity were 0.90 (95% confidence interval: 0.88, 0.92) and 0.72 (95% confidence interval: 0.67, 0.77), respectively. The performance of breast MR imaging was influenced by the prevalence of cancer in the studied population (P = .05) and by whether two criteria (ie, morphology, enhancement, and kinetic enhancement pattern)--versus one or three criteria--were used to differentiate benign from malignant lesions (P = .02).

CONCLUSION

MR imaging of the breast has high sensitivity and lower specificity in the evaluation of breast lesions.

SUPPLEMENTAL MATERIAL

http://radiology.rsnajnls.org/cgi/content/full/2461061298/DC1.

Multiscale analysis of MR-mammography data

In this work we propose a method for automatically discriminating between different types of tissue in MR mammography datasets. This is accomplished by employing a wavelet-based multiscale analysis. After the data has been wavelet-transformed unsupervised machine learning methods are employed to identify typical patterns in the wavelet domain. To demonstrate the potential of the proposed approach we apply a filtering procedure that extracts the wavelet-based image information related to tumour tissue. In this way we obtain a robust segmentation of suspicious tissue in the MR image.

Classification of Signal-Time Curves Obtained by Dynamic Magnetic Resonance Mammography

OBJECTIVE

This study compares the performance of quantitative methods for the characterization of signal-time curves acquired by dynamic contrast-enhanced magnetic resonance mammography from 253 females.

MATERIALS AND METHODS

Signal-time curves obtained from 105 parenchyma, 162 malignant, and 91 benign tissue regions were examined (243 lesions were histopathologically validated). A neural network, a nearest-neighbor, and a threshold classifier were applied to either the entire signal-time curve or pharmacokinetic and descriptive parameters calculated from the curves to differentiate between 2 (malignant or benign) or 3 tissue classes (malignant, benign, or parenchyma). The classifiers were tuned and evaluated according to their performance on 2 distinct subsets of the curves.

RESULTS

The accuracy determined for the neural network and the nearest-neighbor classifiers was nearly identical (approximately 80% in case of 3 tissue classes, and approximately 76% in case of the 2 classes). In contrast, the accuracy of the threshold classifier applied to the discrimination of 3 classes was low (65%).

CONCLUSION

Quantitative classifiers can support the radiologist in the diagnosis of breast lesions.

Neural network approach to the segmentation and classification of dynamic magnetic resonance images of the breast: comparison with empiric and quantitative kinetic parameters

RATIONALE AND OBJECTIVE

An artificial neural network (ANN)-based segmentation method was developed for dynamic contrast-enhanced magnetic resonance (MR) imaging of the breast and compared with quantitative and empiric parameter mapping techniques.

MATERIALS AND METHODS

The study population was composed of 10 patients with seven malignant and three benign lesions undergoing dynamic MR imaging of the breast. All lesions were biopsied or surgically excised, and examined by means of histopathology. A T1-weighted 3D FLASH (fast low angle shot sequence) was acquired before and seven times after the intravenous administration of gadopentetate dimeglumine at a dose of 0.1 mmol/kg body weight. Motion artifacts on MR images were eliminated by voxel-based affine and nonrigid registration techniques. A two-layered feed-forward back-propagation network was created for pixel-by-pixel classification of signal intensity-time curves into benign/malignant tissue types. ANN output was statistically compared with percent-enhancement (E), signal enhancement ratio (SER), time-to-peak, subtracted signal intensity (SUB), pharmacokinetic parameter rate constant (k(ep)), and correlation coefficient to a predefined reference washout curve.

RESULTS

ANN was successfully applied to the classification of breast MR images identifying structures with benign or malignant enhancement kinetics. Correlation coefficient (logistic regression, odds ratio [OR] = 12.9; 95% CI: 7.7-21.8), k(ep) (OR = 1.8; 95% CI: 1.2-2.6), and time-to-peak (OR = 0.45; 95% CI: 0.3-0.7) were independently associated to ANN output classes. SER, E, and SUB were nonsignificant covariates.

CONCLUSION

ANN is capable of classifying breast lesions on MR images. Mapping correlation coefficient, k(ep) and time-to-peak showed the highest association with the ANN result.

Improved artificial neural networks in prediction of malignancy of lesions in contrast-enhanced MR-mammography

The aim of this study was to evaluate the capability of improved artificial neural networks (ANN) and additional novel training methods in distinguishing between benign and malignant breast lesions in contrast-enhanced magnetic resonance-mammography (MRM). A total of 604 histologically proven cases of contrast-enhanced lesions of the female breast at MRI were analyzed. Morphological, dynamic and clinical parameters were collected and stored in a database. The data set was divided into several groups using random or experimental methods [Training & Testing (T&T) algorithm] to train and test different ANNs. An additional novel computer program for input variable selection was applied. Sensitivity and specificity were calculated and compared with a statistical method and an expert radiologist. After optimization of the distribution of cases among the training and testing sets by the T & T algorithm and the reduction of input variables by the Input Selection procedure a highly sophisticated ANN achieved a sensitivity of 93.6% and a specificity of 91.9% in predicting malignancy of lesions within an independent prediction sample set. The best statistical method reached a sensitivity of 90.5% and a specificity of 68.9%. An expert radiologist performed better than the statistical method but worse than the ANN (sensitivity 92.1%, specificity 85.6%). Features extracted out of dynamic contrast-enhanced MRM and additional clinical data can be successfully analyzed by advanced ANNs. The quality of the resulting network strongly depends on the training methods, which are improved by the use of novel training tools. The best results of an improved ANN outperform expert radiologists.

Application of artificial neural networks to the analysis of dynamic MR imaging features of the breast

The discriminative ability of established diagnostic criteria for MRI of the breast is assessed, and their relative relevance using artificial neural networks (ANNs) is determined. A total of 89 women with 105 histopathologically verified breast lesions (73 invasive cancers, 2 in situ cancers, and 30 benign lesions) were included in this study. A T1-weighted 3D FLASH sequence was acquired before and seven times after the intravenous administration of gadopentetate dimeglumine at a dose of 0.2 mmol/kg body weight. ANN models were built to test the discriminative ability of kinetic, morphologic, and combined MR features. The subjects were randomly divided into two parts: a training set of 59 lesions and a verification set of 46 lesions. The training set was used for learning, and the performance of each model was evaluated on the verification set by measuring the area under the ROC curve (Az). An optimally minimized model was constructed using the most relevant input variables that were determined by the automatic relevance determination (ARD) method. ANN models were compared with the performance of a human reader. Margin type, time-to-peak enhancement, and washout ratio showed the highest discriminative ability among diagnostic criteria and comprised the minimized model. Compared with the expert radiologist (Az = 0.799), using the same prediction scale, the minimized ANN model performed best (Az = 0.771), followed by the best kinetic (Az = 0.743), the maximized (Az = 0.727), and the morphologic model (Az = 0.678). The performance of a neural network prediction model is comparable to that of an expert radiologist. A neurostatistical approach is preferred for the analysis of diagnostic criteria when many parameters are involved and complex nonlinear relationships exist in the data set.

Comparison between radiological and artificial neural network diagnosis in clinical screening

The imaging protocol of the UK multicentre magnetic resonance imaging study for screening in women at genetic risk of breast cancer aims to assist in detecting and diagnosing malignant breast lesions. In this paper, we evaluate a three-layer, feed-forward, backpropagation neural network as an artificial radiological classifier using receiver operating characteristic (ROC) curve analysis and compare the results with those obtained using a proposed radiological scoring system for the study which currently supplements the radiologist's clinical opinion, in comparison with histological diagnosis. Based on the 76 symptomatic cases evaluated, descriptive features scored by radiologists showed considerable overlap between benign and malignant, although some features such as irregular contours and heterogeneous enhancement were more often associated with malignant pathology. In this preliminary evaluation, ROC analysis showed that the proposed scoring scheme did not perform well, indicating further refinement is required. When all 23 features were used in the neural network, its performance was poorer than that of the scoring scheme. When only ten features were used, limited to descriptors of enhancement characteristics, the neural network performed similar to the scoring scheme. This comparison shows that the neural network approach to clinical diagnosis has considerable potential and warrants further development.