Automated breast mass detection in 3D reconstructed tomosynthesis volumes: a featureless approach

The purpose of this study was to propose and implement a computer aided detection (CADe) tool for breast tomosynthesis. This task was accomplished in two stages-a highly sensitive mass detector followed by a false positive (FP) reduction stage. Breast tomosynthesis data from 100 human subject cases were used, of which 25 subjects had one or more mass lesions and the rest were normal. For stage 1, filter parameters were optimized via a grid search. The CADe identified suspicious locations were reconstructed to yield 3D CADe volumes of interest. The first stage yielded a maximum sensitivity of 93% with 7.7 FPs/breast volume. Unlike traditional CADe algorithms in which the second stage FP reduction is done via feature extraction and analysis, instead information theory principles were used with mutual information as a similarity metric. Three schemes were proposed, all using leave-one-case-out cross validation sampling. The three schemes, A, B, and C, differed in the composition of their knowledge base of regions of interest (ROIs). Scheme A's knowledge base was comprised of all the mass and FP ROIs generated by the first stage of the algorithm. Scheme B had a knowledge base that contained information from mass ROIs and randomly extracted normal ROIs. Scheme C had information from three sources of information-masses, FPs, and normal ROIs. Also, performance was assessed as a function of the composition of the knowledge base in terms of the number of FP or normal ROIs needed by the system to reach optimal performance. The results indicated that the knowledge base needed no more than 20 times as many FPs and 30 times as many normal ROIs as masses to attain maximal performance. The best overall system performance was 85% sensitivity with 2.4 FPs per breast volume for scheme A, 3.6 FPs per breast volume for scheme B, and 3 FPs per breast volume for scheme C.     

Computer aided detection of clusters of microcalcifications on full field digital mammograms

We are developing a computer-aided detection (CAD) system to identify microcalcification clusters (MCCs) automatically on full field digital mammograms (FFDMs). The CAD system includes six stages: preprocessing; image enhancement; segmentation of microcalcification candidates; false positive (FP) reduction for individual microcalcifications; regional clustering; and FP reduction for clustered microcalcifications. At the stage of FP reduction for individual microcalcifications, a truncated sum-of-squares error function was used to improve the efficiency and robustness of the training of an artificial neural network in our CAD system for FFDMs. At the stage of FP reduction for clustered microcalcifications, morphological features and features derived from the artificial neural network outputs were extracted from each cluster. Stepwise linear discriminant analysis (LDA) was used to select the features. An LDA classifier was then used to differentiate clustered microcalcifications from FPs. A data set of 96 cases with 192 images was collected at the University of Michigan. This data set contained 96 MCCs, of which 28 clusters were proven by biopsy to be malignant and 68 were proven to be benign. The data set was separated into two independent data sets for training and testing of the CAD system in a cross-validation scheme. When one data set was used to train and validate the convolution neural network (CNN) in our CAD system, the other data set was used to evaluate the detection performance. With the use of a truncated error metric, the training of CNN could be accelerated and the classification performance was improved. The CNN in combination with an LDA classifier could substantially reduce FPs with a small tradeoff in sensitivity. By using the free-response receiver operating characteristic methodology, it was found that our CAD system can achieve a cluster-based sensitivity of 70, 80, and 90 % at 0.21, 0.61, and 1.49 FPs/image, respectively. For case-based performance evaluation, a sensitivity of 70, 80, and 90 % can be achieved at 0.07, 0.17, and 0.65 FPs/image, respectively. We also used a data set of 216 mammograms negative for clustered microcalcifications to further estimate the FP rate of our CAD system. The corresponding FP rates were 0.15, 0.31, and 0.86 FPs/image for cluster-based detection when negative mammograms were used for estimation of FP rates.     

Bilateral analysis based false positive reduction for computer-aided mass detection

We have developed a false positive (FP) reduction method based on analysis of bilateral mammograms for computerized mass detection systems. The mass candidates on each view were first detected by our unilateral computer-aided detection (CAD) system. For each detected object, a regional registration technique was used to define a region of interest (ROI) that is "symmetrical" to the object location on the contralateral mammogram. Texture features derived from the spatial gray level dependence matrices and morphological features were extracted from the ROI containing the detected object on a mammogram and its corresponding ROI on the contralateral mammogram. Bilateral features were then generated from corresponding pairs of unilateral features for each object. Two linear discriminant analysis (LDA) classifiers were trained from the unilateral and the bilateral feature spaces, respectively. Finally, the scores from the unilateral LDA classifier and the bilateral LDA asymmetry classifier were fused with a third LDA whose output score was used to distinguish true mass from FPs. A data set of 341 cases of bilateral two-view mammograms was used in this study, of which 276 cases with 552 bilateral pairs contained 110 malignant and 166 benign biopsy-proven masses and 65 cases with 130 bilateral pairs were normal. The mass data set was divided into two subsets for twofold cross-validation training and testing. The normal data set was used for estimation of FP rates. It was found that our bilateral CAD system achieved a case-based sensitivity of 70%, 80%, and 85% at average FP rates of 0.35, 0.75, and 0.95 FPs/image, respectively, on the test data sets with malignant masses. In comparison to the average FP rates for the unilateral CAD system of 0.58, 1.33, and 1.63, respectively, at the corresponding sensitivities, the FP rates were reduced by 40%, 44%, and 42% with the bilateral symmetry information. The improvement was statistically significance (p < 0.05) as estimated by JAFROC analysis.     

False-positive reduction in CAD mass detection using a competitive classification strategy

High false-positive (FP) rate remains to be one of the major problems to be solved in CAD study because too many false-positively cued signals will potentially degrade the performance of detecting true-positive regions and increase the call-back rate in CAD environment. In this paper, we proposed a novel classification method for FP reduction, where the conventional "hard" decision classifier is cascaded with a "soft" decision classification with the objective to reduce false-positives in the cases with multiple FPs retained after the "hard" decision classification. The "soft" classification takes a competitive classification strategy in which only the "best" ones are selected from the pre-classified suspicious regions as the true mass in each case. A neural network structure is designed to implement the proposed competitive classification. Comparative studies of FP reduction on a database of 79 images by a "hard" decision classification and a combined "hard"-"soft" classification method demonstrated the efficiency of the proposed classification strategy. For example, for the high FP sub-database which has only 31.7% of total images but accounts for 63.5% of whole FPs generated in single "hard" classification, the FPs can be reduced for 56% (from 8.36 to 3.72 per image) by using the proposed method at the cost of 1% TP loss (from 69% to 68%) in whole database, while it can only be reduced for 27% (from 8.36 to 6.08 per image) by simply increasing the threshold of "hard" classifier with a cost of TP loss as high as 14% (from 69% to 55%). On the average in whole database, the FP reduction by hybrid "hard"-"soft" classification is 1.58 per image as compared to 1.11 by "hard" classification at the TP costs described above. Because the cases with high dense tissue are of higher risk of cancer incidence and false-negative detection in mammogram screening, and usually generate more FPs in CAD detection, the method proposed in this paper will be very helpful in improving the performance of early detection of breast cancer with CAD.     

Feature-guided analysis for reduction of false positives in CAD of polyps for computed tomographic colonography

We evaluated the effect of our novel technique of feature-guided analysis of polyps on the reduction of false-positive (FP) findings generated by our computer-aided diagnosis (CAD) scheme for the detection of polyps from computed tomography colonographic data sets. The detection performance obtained by use of feature-guided analysis in the segmentation and feature analysis of polyp candidates was compared with that obtained by use of our previously employed fuzzy clustering technique. We also evaluated the effect of a feature called modified gradient concentration (MGC) on the detection performance. A total of 144 data sets, representing prone and supine views of 72 patients that included 14 patients with 21 colorectal polyps 5-25 mm in diameter, were used in the evaluation. At a 100% by-patient (95% by-polyp) detection sensitivity, the FP rate of our CAD scheme with feature-guided analysis based on round-robin evaluation was 1.3 (1.5) FP detections per patient. This corresponds to a 70-75% reduction in the number of FPs obtained by use of fuzzy clustering at the same sensitivity levels. Application of the MGC feature instead of our previously used gradient concentration feature did not improve the detection result. The results indicate that feature-guided analysis is useful for achieving high sensitivity and a low FP rate in our CAD scheme.     

Joint two-view information for computerized detection of microcalcifications on mammograms

We are developing new techniques to improve the accuracy of computerized microcalcification detection by using the joint two-view information on craniocaudal (CC) and mediolateral-oblique (MLO) views. After cluster candidates were detected using a single-view detection technique, candidates on CC and MLO views were paired using their radial distances from the nipple. Candidate pairs were classified with a similarity classifier that used the joint information from both views. Each cluster candidate was also characterized by its single-view features. The outputs of the similarity classifier and the single-view classifier were fused and the cluster candidate was classified as a true microcalcification cluster or a false-positive (FP) using the fused two-view information. A data set of 116 pairs of mammograms containing microcalcification clusters and 203 pairs of normal images from the University of South Florida (USF) public database was used for training the two-view detection algorithm. The trained method was tested on an independent test set of 167 pairs of mammograms, which contained 71 normal pairs and 96 pairs with microcalcification clusters collected at the University of Michigan (UM). The similarity classifier had a very low FP rate for the test set at low and medium levels of sensitivity. However, the highest mammogram-based sensitivity that could be reached by the similarity classifier was 69%. The single-view classifier had a higher FP rate compared to the similarity classifier, but it could reach a maximum mammogram-based sensitivity of 93%. The fusion method combined the scores of these two classifiers so that the number of FPs was substantially reduced at relatively low and medium sensitivities, and a relatively high maximum sensitivity was maintained. For the malignant microcalcification clusters, at a mammogram-based sensitivity of 80%, the FP rates were 0.18 and 0.35 with the two-view fusion and single-view detection methods, respectively. When the training and test sets were switched, a similar improvement was obtained, except that both the fusion and single-view detection methods had superior test performances on the USF data set than those on the UM data set. Our results indicate that correspondence of cluster candidates on two different views provides valuable additional information for distinguishing FPs from true microcalcification clusters.     

An SVM classifier to separate false signals from microcalcifications in digital mammograms.

In this paper we investigate the feasibility of using an SVM (support vector machine) classifier in our automatic system for the detection of clustered microcalcifications in digital mammograms. SVM is a technique for pattern recognition which relies on the statistical learning theory. It minimizes a function of two terms: the number of misclassified vectors of the training set and a term regarding the generalization classifier capability. We compare the SVM classifier with an MLP (multi-layer perceptron) in the false-positive reduction phase of our detection scheme: a detected signal is considered either microcalcification or false signal, according to the value of a set of its features. The SVM classifier gets slightly better results than the MLP one (Az value of 0.963 against 0.958) in the presence of a high number of training data; the improvement becomes much more evident (Az value of 0.952 against 0.918) in training sets of reduced size. Finally, the setting of the SVM classifier is much easier than the MLP one.     

EEG Feature Pre-processing for Neonatal Epileptic Seizure Detection

Aim of our project is to further optimize neonatal seizure detection using support vector machine (SVM). First, a Kalman filter (KF) was used to filter both feature and classifier output time series in order to increase temporal precision. Second, EEG baseline feature correction (FBC) was introduced to reduce inter patient variability in feature distributions. The performance of the detection methods is evaluated on 54 multi channel routine EEG recordings from 39 both term and pre-term newborns. The area under the receiver operating characteristics curve (AUC) as well as sensitivity and specificity are used to evaluate the performance of the classification method. SVM without KF and FBC achieves an AUC of 0.767 (sensitivity 0.679, specificity 0.707). The highest AUC of 0.902 (sensitivity 0.801, specificity 0.831) is achieved on baseline corrected features with a Kalman smoother used for training data pre-processing and a KF used to filter the classifier output. Both FBC and KF significantly improve neonatal epileptic seizure detection. This paper introduces significant improvements for the state of the art SVM based neonatal epileptic seizure detection.     

Combined adaptive enhancement and region-growing segmentation of breast masses on digitized mammograms.

As an ongoing effort to develop a computer aid for detection of masses on mammograms, we recently designed an object-based region-growing technique to improve mass segmentation. This segmentation method utilizes the density-weighted contrast enhancement (DWCE) filter as a pre-processing step. The DWCE filter adaptively enhances the contrast between the breast structures and the background. Object-based region growing was then applied to each of the identified structures. The region-growing technique uses gray-scale and gradient information to adjust the initial object borders and to reduce merging between adjacent or overlapping structures. Each object is then classified as a breast mass or normal tissue based on extracted morphological and texture features. In this study we evaluated the sensitivity of this combined segmentation scheme and its ability to reduce false positive (FP) detections on a data set of 253 digitized mammograms, each of which contained a biopsy-proven breast mass. It was found that the segmentation scheme detected 98% of the 253 biopsy-proven breast masses in our data set. After final FP reduction, the detection resulted in 4.2 FP per image at a 90% true positive (TP) fraction and 2.0 FPs per image at an 80% TP fraction. The combined DWCE and object-based region growing technique increased the initial detection sensitivity, reduced merging between neighboring structures, and reduced the number of FP detections in our automated breast mass detection scheme.     

Computer-aided detection of masses in digital tomosynthesis mammography: comparison of three approaches

The authors are developing a computer-aided detection (CAD) system for masses on digital breast tomosynthesis mammograms (DBT). Three approaches were evaluated in this study. In the first approach, mass candidate identification and feature analysis are performed in the reconstructed three-dimensional (3D) DBT volume. A mass likelihood score is estimated for each mass candidate using a linear discriminant analysis (LDA) classifier. Mass detection is determined by a decision threshold applied to the mass likelihood score. A free response receiver operating characteristic (FROC) curve that describes the detection sensitivity as a function of the number of false positives (FPs) per breast is generated by varying the decision threshold over a range. In the second approach, prescreening of mass candidate and feature analysis are first performed on the individual two-dimensional (2D) projection view (PV) images. A mass likelihood score is estimated for each mass candidate using an LDA classifier trained for the 2D features. The mass likelihood images derived from the PVs are backprojected to the breast volume to estimate the 3D spatial distribution of the mass likelihood scores. The FROC curve for mass detection can again be generated by varying the decision threshold on the 3D mass likelihood scores merged by backprojection. In the third approach, the mass likelihood scores estimated by the 3D and 2D approaches, described above, at the corresponding 3D location are combined and evaluated using FROC analysis. A data set of 100 DBT cases acquired with a GE prototype system at the Breast Imaging Laboratory in the Massachusetts General Hospital was used for comparison of the three approaches. The LDA classifiers with stepwise feature selection were designed with leave-one-case-out resampling. In FROC analysis, the CAD system for detection in the DBT volume alone achieved test sensitivities of 80% and 90% at average FP rates of 1.94 and 3.40 per breast, respectively. With the 2D detection approach, the FP rates were 2.86 and 4.05 per breast, respectively, at the corresponding sensitivities. In comparison, the average FP rates of the system combining the 3D and 2D information were 1.23 and 2.04 per breast, respectively, at 80% and 90% sensitivities. The difference in the detection performances between the 2D and the 3D approach, and that between the 3D and the combined approach were both statistically significant (p = 0.02 and 0.01, respectively) as estimated by alternative FROC analysis. The combined system is a promising approach to improving automated mass detection on DBTs.     

A comparison approach toward finding the best feature and classifier in cue-based BCI

In this paper, a comparative evaluation of state-of-the art feature extraction and classification methods is presented for five subjects in order to increase the performance of a cue-based Brain-Computer interface (BCI) system for imagery tasks (left and right hand movements). To select an informative feature with a reliable classifier features containing standard bandpower, AAR coefficients, and fractal dimension along with support vector machine (SVM), Adaboost and Fisher linear discriminant analysis (FLDA) classifiers have been assessed. In the single feature-classifier combinations, bandpower with FLDA gave the best results for three subjects, and fractal dimension and FLDA and SVM classifiers lead to the best results for two other subjects. A genetic algorithm has been used to find the best combination of the features with the aforementioned classifiers and led to dramatic reduction of the classification error and also best results in the four subjects. Genetic feature combination results have been compared with the simple feature combination to show the performance of the Genetic algorithm.     

A novel approach for detection and classification of mammographic microcalcifications using wavelet analysis and extreme learning machine

The objective of this paper is to reveal the effectiveness of wavelet based tissue texture analysis for microcalcification detection in digitized mammograms using Extreme Learning Machine (ELM). Microcalcifications are tiny deposits of calcium in the breast tissue which are potential indicators for early detection of breast cancer. The dense nature of the breast tissue and the poor contrast of the mammogram image prohibit the effectiveness in identifying microcalcifications. Hence, a new approach to discriminate the microcalcifications from the normal tissue is done using wavelet features and is compared with different feature vectors extracted using Gray Level Spatial Dependence Matrix (GLSDM) and Gabor filter based techniques. A total of 120 Region of Interests (ROIs) extracted from 55 mammogram images of mini-Mias database, including normal and microcalcification images are used in the current research. The network is trained with the above mentioned features and the results denote that ELM produces relatively better classification accuracy (94%) with a significant reduction in training time than the other artificial neural networks like Bayesnet classifier, Naivebayes classifier, and Support Vector Machine. ELM also avoids problems like local minima, improper learning rate, and over fitting.     

Classification of breast masses in mammograms using genetic programming and feature selection

Mammography is a widely used screening tool and is the gold standard for the early detection of breast cancer. The classification of breast masses into the benign and malignant categories is an important problem in the area of computer-aided diagnosis of breast cancer. A small dataset of 57 breast mass images, each with 22 features computed, was used in this investigation; the same dataset has been previously used in other studies. The extracted features relate to edge-sharpness, shape, and texture. The novelty of this paper is the adaptation and application of the classification technique called genetic programming (GP), which possesses feature selection implicitly. To refine the pool of features available to the GP classifier, we used feature-selection methods, including the introduction of three statistical measures—Student’s t test, Kolmogorov–Smirnov test, and Kullback–Leibler divergence. Both the training and test accuracies obtained were high: above 99.5% for training and typically above 98% for test experiments. A leave-one-out experiment showed 97.3% success in the classification of benign masses and 95.0% success in the classification of malignant tumors. A shape feature known as fractional concavity was found to be the most important among those tested, since it was automatically selected by the GP classifier in almost every experiment.     

High efficient system for automatic classification of the electrocardiogram beats

Automatic classification of the electrocardiogram (ECG) signals is an important subject for clinical diagnosis of heart disease. This study investigates the design of a high-efficient system to classify five types of ECG beat namely normal beats and four manifestations of heart arrhythmia, in twofold. First, we propose a system that includes two main modules: a feature extraction module and a classification module. Feature extraction module extracts a suitable combination of the ECG’s morphological characteristics and timing interval features. Discrete wavelet transform is used to extract the morphological features. In the classification module, a multi-class support vector machine (SVM)-based classifier is employed. The parameters of this system are determined based on a trial and error method and its performance is evaluated for the MIT-BIH arrhythmia database. Extensive experiments on the parameters of this system such as classifier kernels and various types of features are conducted. These experiments show that in SVM training, the kernels, kernel parameters, and feature selection have very important roles for SVM classification accuracy. Therefore, most appropriates of these parameters should be used for SVM training. Then at the second fold, a novel hybrid intelligent system (HIS) is proposed that consists of three main modules. In the HIS, further to the two mentioned modules, an optimization module is added. In this module, a genetic algorithm is used for optimization of the relevant parameters of system. These parameters are: wavelet filter type for feature extraction, wavelet decomposition level, and classifier’s parameters. Experimental results show that optimization improves the recognition system, efficiently, and HIS is more superior to the system, which as constant parameters.     

Massive-training artificial neural network (MTANN) for reduction of false positives in computer-aided detection of polyps: Suppression of rectal tubes

One of the limitations of the current computer-aided detection (CAD) of polyps in CT colonography (CTC) is a relatively large number of false-positive (FP) detections. Rectal tubes (RTs) are one of the typical sources of FPs because a portion of a RT, especially a portion of a bulbous tip, often exhibits a cap-like shape that closely mimics the appearance of a small polyp. Radiologists can easily recognize and dismiss RT-induced FPs; thus, they may lose their confidence in CAD as an effective tool if the CAD scheme generates such "obvious" FPs due to RTs consistently. In addition, RT-induced FPs may distract radiologists from less common true positives in the rectum. Therefore, removal RT-induced FPs as well as other types of FPs is desirable while maintaining a high sensitivity in the detection of polyps. We developed a three-dimensional (3D) massive-training artificial neural network (MTANN) for distinction between polyps and RTs in 3D CTC volumetric data. The 3D MTANN is a supervised volume-processing technique which is trained with input CTC volumes and the corresponding "teaching" volumes. The teaching volume for a polyp contains a 3D Gaussian distribution, and that for a RT contains zeros for enhancement of polyps and suppression of RTs, respectively. For distinction between polyps and nonpolyps including RTs, a 3D scoring method based on a 3D Gaussian weighting function is applied to the output of the trained 3D MTANN. Our database consisted of CTC examinations of 73 patients, scanned in both supine and prone positions (146 CTC data sets in total), with optical colonoscopy as a reference standard for the presence of polyps. Fifteen patients had 28 polyps, 15 of which were 5-9 mm and 13 were 10-25 mm in size. These CTC cases were subjected to our previously reported CAD scheme that included centerline-based segmentation of the colon, shape-based detection of polyps, and reduction of FPs by use of a Bayesian neural network based on geometric and texture features. Application of this CAD scheme yielded 96.4% (27/28) by-polyp sensitivity with 3.1 (224/73) FPs per patient, among which 20 FPs were caused by RTs. To eliminate the FPs due to RTs and possibly other normal structures, we trained a 3D MTANN with ten representative polyps and ten RTs, and applied the trained 3D MTANN to the above CAD true- and false-positive detections. In the output volumes of the 3D MTANN, polyps were represented by distributions of bright voxels, whereas RTs and other normal structures partly similar to RTs appeared as darker voxels, indicating the ability of the 3D MTANN to suppress RTs as well as other normal structures effectively. Application of the 3D MTANN to the CAD detections showed that the 3D MTANN eliminated all RT-induced 20 FPs, as well as 53 FPs due to other causes, without removal of any true positives. Overall, the 3D MTANN was able to reduce the FP rate of the CAD scheme from 3.1 to 2.1 FPs per patient (33% reduction), while the original by-polyp sensitivity of 96.4% was maintained.     

基于机器学习的心音识别分类研究

心音信号可反映心脏的病理信息,是诊断心脏健康的重要依据之一。本文首先从心音信号提取时频域、梅尔倒谱系数等145个特征作为机器学习的输入数据集,然后在随机森林、LightGBM、XGBoost、GBDT、SVM共5种分类器中选出效果最佳分类器与递归特征消除算法结合进行数据挖掘,找出重要特征集并对其分类效果做比较与分析,最后运用Stacking模型融合方法优化模型。数据挖掘特征子集比同数量特征子集在准确率、召回率、精确率、F1值上分别提高了33.51%、14.54%、20.61%、24.04%;采用LightGBM和SVM模型融合可将F1值提高至92.6%。本文提出了一种有效的心音识别分类方法,挖掘出心音最重要的8个特征,为临床诊断提供参考。     

Tolerance to missing data using a likelihood ratio based classifier for computer-aided classification of breast cancer

While mammography is a highly sensitive method for detecting breast tumours, its ability to differentiate between malignant and benign lesions is low, which may result in as many as 70% of unnecessary biopsies. The purpose of this study was to develop a highly specific computer-aided diagnosis algorithm to improve classification of mammographic masses. A classifier based on the likelihood ratio was developed to accommodate cases with missing data. Data for development included 671 biopsy cases (245 malignant), with biopsy-proved outcome. Sixteen features based on the BI-RADS lexicon and patient history had been recorded for the cases, with 1.3 +/- 1.1 missing feature values per case. Classifier evaluation methods included receiver operating characteristic and leave-one-out bootstrap sampling. The classifier achieved 32% specificity at 100% sensitivity on the 671 cases with 16 features that had missing values. Utilizing just the seven features present for all cases resulted in decreased performance at 100% sensitivity with average 19% specificity. No cases and no feature data were omitted during classifier development, showing that it is more beneficial to utilize cases with missing values than to discard incomplete cases that cannot be handled by many algorithms. Classification of mammographic masses was commendable at high sensitivity levels, indicating that benign cases could be potentially spared from biopsy.     

GMDH-based feature ranking and selection for improved classification of medical data

Medical applications are often characterized by a large number of disease markers and a relatively small number of data records. We demonstrate that complete feature ranking followed by selection can lead to appreciable reductions in data dimensionality, with significant improvements in the implementation and performance of classifiers for medical diagnosis. We describe a novel approach for ranking all features according to their predictive quality using properties unique to learning algorithms based on the group method of data handling (GMDH). An abductive network training algorithm is repeatedly used to select groups of optimum predictors from the feature set at gradually increasing levels of model complexity specified by the user. Groups selected earlier are better predictors. The process is then repeated to rank features within individual groups. The resulting full feature ranking can be used to determine the optimum feature subset by starting at the top of the list and progressively including more features until the classification error rate on an out-of-sample evaluation set starts to increase due to overfitting. The approach is demonstrated on two medical diagnosis datasets (breast cancer and heart disease) and comparisons are made with other feature ranking and selection methods. Receiver operating characteristics (ROC) analysis is used to compare classifier performance. At default model complexity, dimensionality reduction of 22 and 54% could be achieved for the breast cancer and heart disease data, respectively, leading to improvements in the overall classification performance. For both datasets, considerable dimensionality reduction introduced no significant reduction in the area under the ROC curve. GMDH-based feature selection results have also proved effective with neural network classifiers.