Detection of lung nodules in digital chest radiographs using artificial neural networks: a pilot study

2D Statistical Lung Shape Analysis Using Chest Radiographs: Modelling and Segmentation

Accurate information of the lung shape analysis and its anatomical variations is very noticeable in medical imaging. The normal variations of the lung shape can be interpreted as a normal lung. In contrast, abnormal variations of the lung shape can be a result of one of the pulmonary diseases. The goal of this study is twofold: (1) represent two lung shape models which are different at the reference points in registration process considering to show their impact on estimating the inter-patient 2D lung shape variations and (2) using the obtained models in lung field segmentation by utilizing active shape model (ASM) technique. The represented models which showed the inter-patient 2D lung shape variations in two different forms are fully compared and evaluated. The results show that the models along with standard principal component analysis (PCA) can be able to explain more than 95% of total variations in all cases using only first 7 principal component (PC) modes for both lungs. Both models are used in ASM-based segmentation technique for lung field segmentation. The segmentation results are evaluated using leave-one-out cross validation technique. According to the experimental results, the proposed method has average dice similarity coefficient of 97.1% and 96.1% for the right and the left lung, respectively. The results show that the proposed segmentation method is more stable and accurate than other model-based techniques to inter-patient lung field segmentation.

ADAPTIVE SIMULATED ANNEALING FOR CT IMAGE CLASSIFICATION

We present a pattern classification method that combines the classical Perceptron algorithm with simulated annealing. For a sample set S of n-dimensional patterns labeled as positive and negative, our algorithm computes threshold circuits of small depth where the linear threshold functions of the first layer are calculated by simulated annealing with the logarithmic cooling schedule c(k) = Γ(k)/ ln (k + 2). The parameter Γ depends on the sample set and changes in time, and the neighborhood relation is determined by the Perceptron algorithm. We apply the approach to the recognition of focal liver tumours. From 400 positive (focal liver tumour) and 400 negative (normal liver tissue) examples a depth-six threshold circuit is calculated. The examples are of size n = 14161 = 119 × 119 and they are presented in the DICOM format. On test sets of 100 + 100 examples (disjoint from the learning set) we obtain a correct classification of more than 98%.

Artificial neural network: border detection in echocardiography

Being non-invasive and low cost, the echocardiography has become a diagnostic technique largely applied for the determination of the left ventricle systolic and diastolic volumes, which are used indirectly to calculate the left ventricle ejection volume, the cardiac cavities muscular contraction, the regional ejection fraction, the myocardial thickness, and the ventricular mass, etc. However, the image is very noisy, which renders the delineation of the borders of the left ventricle very difficult. While there are many techniques image segmentation, this work chooses the artificial neural network (ANN) since it is not very sensitive to noise. In order to reduce the processing time, the operator selects the region of interest where the neural network will identify the borders. Neighborhood and gradient search techniques are then employed to link the points and the left ventricle contour is traced. The present method has been efficient in detecting the left ventricle borders echocardiography images compared to those whose borders were delineated by the specialists. For good results, it is important to choose properly the areas to be analyzed and the central points of these areas.

Bounded-depth threshold circuits for computer-assisted CT image classification

We present a stochastic algorithm that computes threshold circuits designed to discriminate between two classes of computed tomography (CT) images. The algorithm employs a partition of training examples into several classes according to the average grey scale value of images. For each class, a sub-circuit is computed, where the first layer of the sub-circuit is calculated by a new combination of the Perceptron algorithm with a special type of simulated annealing. The algorithm is evaluated for the case of liver tissue classification. A depth-five threshold circuit (with pre-processing: depth-seven) is calculated from 400 positive (abnormal findings) and 400 negative (normal liver tissue) examples. The examples are of size n=14,161 (119 x 119) with an 8 bit grey scale. On test sets of 100 positive and 100 negative examples (all different from the learning set) we obtain a correct classification close to 99%. The total sequential run-time to compute a depth-five circuit is about 75h up to 230h on a SUN Ultra 5/360 workstation, depending on the width of the threshold circuit at depth-three. In our computational experiments, the depth-five circuits were calculated from three simultaneous runs for depth-four circuits. The classification of a single image is performed within a few seconds.

Computerized classification of suspicious regions in chest radiographs using subregion Hotelling observers

We propose to investigate the use of subregion Hotelling observers (SRHOs) in conjunction with perceptrons for the computerized classification of suspicious regions in chest radiographs for being nodules requiring follow up. Previously, 239 regions of interest (ROIs), each containing a suspicious lesion with proven classification, were collected. We chose to investigate the use of SRHOs as part of a multilayer classifier to determine the presence of a nodule. Each SRHO incorporates information about signal, background, and noise correlation for classification. For this study, 225 separate Hotelling observers were set up in a grid across each ROI. Each separate observer discriminates an 8 by 8 pixel area. A round robin sampling scheme was used to generate the 225 features, where each feature is the output of the individual observers. These features were then rank ordered by the magnitude of the weights of a perceptron. Once rank ordered, subsets of increasing number of features were selected to be used in another perceptron. This perceptron was trained to minimize mean squared error and the output was a continuous variable representing the likelihood of the region being a nodule. Performance was evaluated by receiver operating characteristic (ROC) analysis and reported as the area under the curve (Az). The classifier was optimized by adding additional features until the Az declined. The optimized subset of observers then were combined using a third perceptron. A subset of 80 features was selected which gave an Az of 0.972. Additionally, at 98.6% sensitivity, the classifier had a specificity of 71.3% and increased the positive predictive value from 60.7% to 84.1 %. Preliminary results suggest that using SRHOs in combination with perceptrons can provide a successful classification scheme for pulmonary nodules. This approach could be incorporated into a larger computer aided detection system for decreasing false positives.

Logarithmic simulated annealing for X-ray diagnosis

We present a new stochastic learning algorithm and first results of computational experiments on fragments of liver CT images. The algorithm is designed to compute a depth-three threshold circuit, where the first layer is calculated by an extension of the Perceptron algorithm by a special type of simulated annealing. The fragments of CT images are of size 119x119 with eight bit grey levels. From 348 positive (focal liver tumours) and 348 negative examples a number of hypotheses of the type w(1)x(1)+. . .;+w(n)x(n)>/=theta were calculated for n=14161. The threshold functions at levels two and three were determined by computational experiments. The circuit was tested on various sets of 50+50 additional positive and negative examples. For depth-three circuits, we obtained a correct classification of about 97%. The input to the algorithm is derived from the DICOM standard representation of CT images. The simulated annealing procedure employs a logarithmic cooling schedule c(k)=Gamma/ln(k+2), where Gamma is a parameter that depends on the underlying configuration space. In our experiments, the parameter Gamma is chosen according to estimations of the maximum escape depth from local minima of the associated energy landscape.

Automatic detection and quantification of ground-glass opacities on high-resolution CT using multiple neural networks: comparison with a density mask

OBJECTIVE

We compared multiple neural networks with a density mask for the automatic detection and quantification of ground-glass opacities on high-resolution CT under clinical conditions.

SUBJECTS AND METHODS

Eighty-four patients (54 men and 30 women; age range, 18-82 years; mean age, 49 years) with a total of 99 consecutive high-resolution CT scans were enrolled in the study. The neural network was designed to detect ground-glass opacities with high sensitivity and to omit air-tissue interfaces to increase specificity. The results of the neural network were compared with those of a density mask (thresholds, -750/-300 H), with a radiologist serving as the gold standard.

RESULTS

The neural network classified 6% of the total lung area as ground-glass opacities. The density mask failed to detect 1.3%, and this percentage represented the increase in sensitivity that was achieved by the neural network. The density mask identified another 17.3% of the total lung area to be ground-glass opacities that were not detected by the neural network. This area represented the increase in specificity achieved by the neural network. Related to the extent of the ground-glass opacities as classified by the radiologist, the neural network (density mask) reached a sensitivity of 99% (89%), specificity of 83% (55%), positive predictive value of 78% (18%), negative predictive value of 99% (98%), and accuracy of 89% (58%).

CONCLUSION

Automatic segmentation and quantification of ground-glass opacities on high-resolution CT by a neural network are sufficiently accurate to be implemented for the preinterpretation of images in a clinical environment; it is superior to a double-threshold density mask.

A neural network to predict symptomatic lung injury

A nonlinear neural network that simultaneously uses pre-radiotherapy (RT) biological and physical data was developed to predict symptomatic lung injury. The input data were pre-RT pulmonary function, three-dimensional treatment plan doses and demographics. The output was a single value between 0 (asymptomatic) and 1 (symptomatic) to predict the likelihood that a particular patient would become symptomatic. The network was trained on data from 97 patients for 400 iterations with the goal to minimize the mean-squared error. Statistical analysis was performed on the resulting network to determine the model's accuracy. Results from the neural network were compared with those given by traditional linear discriminate analysis and the dose-volume histogram reduction (DVHR) scheme of Kutcher. Receiver-operator characteristic (ROC) analysis was performed on the resulting network which had Az = 0.833 +/- 0.04. (Az is the area under the ROC curve.) Linear discriminate multivariate analysis yielded an Az = 0.813 +/- 0.06. The DVHR method had Az = 0.521 +/- 0.08. The network was also used to rank the significance of the input variables. Future studies will be conducted to improve network accuracy and to include functional imaging data.

Characteristics of regions suspicious for pulmonary nodules at chest radiography

RATIONALE AND OBJECTIVES

This study was performed to determine physical characteristics of areas on chest radiographs that are suspicious but not definitive for the presence of a pulmonary nodule and the characteristics of areas that contain an obvious nodule.

MATERIALS AND METHODS

Two groups of patients were identified: those who had an area at plain radiography that was suspicious for a pulmonary nodule and underwent fluoroscopy for further evaluation (138 patients, 142 areas) and those who had an obvious nodule at plain radiography who underwent computed tomography for further evaluation (72 patients, 97 areas). The measured characteristics of the region of interest included size, circularity, compactness, contrast, and location.

RESULTS

A comparison of the data show that while there was some difference between these groups of patients with regard to location of the nodules, there were essentially no differences with regard to size, circularity, compactness, and contrast of the regions of interest.

CONCLUSION

Size, circularity, compactness, contrast, and location are not sufficient to distinguish pulmonary nodules from other suspicious regions on the chest radiograph.

Neural network analysis of combined conventional and experimental prognostic markers in prostate cancer: a pilot study

Prostate cancer is the second most common malignancy in men in the UK. The disease is unpredictable in its behaviour and, at present, no single investigative method allows clinicians to differentiate between tumours that will progress and those that will remain quiescent. There is an increasing need for novel means to predict prognosis and outcome of the disease. The aim of this study was to assess the value of artificial neural networks in predicting outcome in prostate cancer in comparison with statistical methods, using a combination of conventional and experimental biological markers. Forty-one patients with different stages and grades of prostate cancer undergoing a variety of treatments were analysed. Artificial neural networks were used as follows: eight input neurons consisting of six conventional factors (age, stage, bone scan findings, grade, serum PSA, treatment) and two experimental markers (immunostaining for bcl-2 and p53, which are both apoptosis-regulating genes). Twenty-one patients were used for training and 20 for testing. A total of 80% of the patients were correctly classified regarding outcome using the combination of factors. When both bcl-2 and p53 immunoreactivity were excluded from the analysis, correct prediction of the outcome was achieved in only 60% of the patients (P = 0.0032). This study was able to demonstrate the value of artificial neural networks in the analysis of prognostic markers in prostate cancer. In addition, the potential for using this technology to evaluate novel markers is highlighted. Further large-scale analyses are required to incorporate this methodology into routine clinical practice.

Fragmentary window filtering for multiscale lung nodule detection: preliminary study

RATIONALE AND OBJECTIVES

The authors evaluated computer-assisted diagnostic (CAD) methods used to detect suspicious areas on lung radiographs.

MATERIALS AND METHODS

The authors designed a fragmentary window filtering (FWF) algorithm for detecting lung nodule patterns, which generally appear as circular areas of high opacity on the chest radiograph. The FWF algorithm helps differentiate circular patterns from overlapping radiographic background. A multiscale analysis was performed to locate multiscale nodules. Receiver operating characteristic analysis was performed by using a lung nodule that was extracted from a chest radiograph. The nodule underwent scalings and subsequent superimposition onto 140 normal regions of interest from six chest radiographs.

RESULTS

The FWF method was superior to the matched filtering method in the detection of suspicious areas.

CONCLUSION

The proposed FWF-based method should provide improved detection of lung nodules on chest radiographs.

Adequacy testing of training set sample sizes in the development of a computer-assisted diagnosis scheme

RATIONALE AND OBJECTIVES

The authors assessed the performance changes of a computer-assisted diagnosis (CAD) scheme as a function of the number of regions used for training (rule-setting).

MATERIALS AND METHODS

One hundred twenty regions depicting actual masses and 400 suspicious but actually negative regions were selected as a testing data set from a database of 2,146 regions identified as suspicious on 618 mammograms. An artificial neural network using 24 and 16 region-based features as input neurons was applied to classify the regions as positive or negative for the presence of a mass. CAD scheme performance was evaluated on the testing data set as the number of regions used for training increased from 60 to 496.

RESULTS

As the number of regions in the training sets increased, the results decreased and plateaued beyond a sample size of approximately 200 regions. Performance with the testing data set continued to improve as the training data set increased in size.

CONCLUSION

A trend in a system's performance as a function of training set size can be used to assess adequacy of the training data set in the development of a CAD scheme.