Artificial convolution neural network for medical image pattern recognition

Selection of an optimal neural network architecture for computer-aided detection of microcalcifications--comparison of automated optimization techniques

Many computer-aided diagnosis (CAD) systems use neural networks (NNs) for either detection or classification of abnormalities. Currently, most NNs are "optimized" by manual search in a very limited parameter space. In this work, we evaluated the use of automated optimization methods for selecting an optimal convolution neural network (CNN) architecture. Three automated methods, the steepest descent (SD), the simulated annealing (SA), and the genetic algorithm (GA), were compared. We used as an example the CNN that classifies true and false microcalcifications detected on digitized mammograms by a prescreening algorithm. Four parameters of the CNN architecture were considered for optimization, the numbers of node groups and the filter kernel sizes in the first and second hidden layers, resulting in a search space of 432 possible architectures. The area Az under the receiver operating characteristic (ROC) curve was used to design a cost function. The SA experiments were conducted with four different annealing schedules. Three different parent selection methods were compared for the GA experiments. An available data set was split into two groups with approximately equal number of samples. By using the two groups alternately for training and testing, two different cost surfaces were evaluated. For the first cost surface, the SD method was trapped in a local minimum 91% (392/432) of the time. The SA using the Boltzman schedule selected the best architecture after evaluating, on average, 167 architectures. The GA achieved its best performance with linearly scaled roulette-wheel parent selection; however, it evaluated 391 different architectures, on average, to find the best one. The second cost surface contained no local minimum. For this surface, a simple SD algorithm could quickly find the global minimum, but the SA with the very fast reannealing schedule was still the most efficient. The same SA scheme, however, was trapped in a local minimum on the first cost surface. Our CNN study demonstrated that, if optimization is to be performed on a cost surface whose characteristics are not known a priori, it is advisable that a moderately fast algorithm such as a SA using a Boltzman cooling schedule be used to conduct an efficient and thorough search, which may offer a better chance of reaching the global minimum.

A computer-aided design mammography screening system for detection and classification of microcalcifications

This paper presents a prototype of a computer-aided design (CAD) diagnostic system for mammography screening to automatically detect and classify microcalcifications (MCCs) in mammograms. It comprises four modules. The first module, called the Mammogram Preprocessing Module, inputs and digitizes mammograms into 8-bit images of size 2048x2048, extracts the breast region from the background, enhances the extracted breast and stores the processed mammograms in a data base. Since only clustered MCCs are of interest in providing a sign of breast cancer, the second module, called the MCCs Finder Module, finds and locates suspicious areas of clustered MCCs, called regions of interest (ROIs). The third module, called the MCCs Detection Module, is a real time computer automated MCCs detection system that takes as inputs the ROIs provided by the MCCs Finder Module. It uses two different window sizes to automatically extract the microcalcifications from the ROIs. It begins with a large window of size 64x64 to quickly screen mammograms to find large calcified areas, this is followed by a smaller window of size 8x8 to extract tiny, isolated microcalcifications. Finally, the fourth module, called the MCCs Classification Module, classifies the detected clustered microcalcifications into five categories according to BI-RADS (Breast Imaging Reporting and Data System) format recommended by the American College of Radiology. One advantage of the designed system is that each module is a separate component that can be individually upgraded to improve the whole system. Despite that it is still is a prototype system a preliminary clinical evaluation at TaiChung Veterans General Hospital (TCVGH) has shown that the system is very flexible and can be integrated with the existing Picture Archiving and Communications System (PACS) currently implemented in the Department of Radiology at TCVGH.

Computerized detection of malignant tumors on digital mammograms

This paper presents a tumor detection system for fully digital mammography. The processing scheme adopted in the proposed system focuses on the solution of two problems. One is how to detect tumors as suspicious regions with a very weak contrast to their background and another is how to extract features which characterize malignant tumors. For the first problem, a unique adaptive filter called the iris filter is proposed. It is very effective in enhancing approximately rounded opacities no matter what their contrasts might be. Clues for differentiation between malignant tumors and other tumors are believed to be mostly in their border areas. This paper proposes typical parameters which reflect boundary characteristics. To confirm the system performance for unknown samples, large scale experiments using 1212 CR images were performed. The results showed that the sensitivity of the proposed system was 90.5% and the average number of false positives per image was found to be only 1.3. These results show the effectiveness of the proposed system.

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.

Optimally weighted wavelet transform based on supervised training for detection of microcalcifications in digital mammograms

We are developing a computer-aided diagnosis (CAD) scheme for detection of clustered microcalcifications in digital mammograms. The use of an empirically chosen wavelet and scale combination for detection of microcalcifications as an initial step of the CAD scheme has been reported by us previously. In this study, we developed a technique for optimizing the weights at individual scales in the wavelet transform to improve the performance of our CAD scheme based on the supervised learning method. In the learning process, an error function was formulated to represent the difference between a desired output and the reconstructed image obtained from weighted wavelet coefficients for a given mammogram. The error function was then minimized by modifying the weights for wavelet coefficients by means of a conjugate gradient algorithm. The Least Asymmetric Daubechies' wavelets were optimized with 297 regions of interest (ROIs) as a training set by a jackknife method. The performance of the optimally weighted wavelets was evaluated by means of receiver-operating characteristic (ROC) analysis by use of the above set of ROIs. The analysis yielded an average area under the ROC curve of 0.92, which outperforms the difference-image technique used in our existing CAD scheme, as well as the partial reconstruction method used in our previous study.

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.

ARTMAP-IC and medical diagnosis: Instance counting and inconsistent cases

For complex database prediction problems such as medical diagnosis, the ARTMAP-IC neural network adds distributed prediction and category instance counting to the basic fuzzy ARTMAP system. For the ARTMAP match tracking algorithm, which controls search following a predictive error, a new version facilitates prediction with sparse or inconsistent data. Compared to the original match tracking algorithm (MT+), the new algorithm (MT-) better approximates the real-time network differential equations and further compresses memory without loss of performance. Simulations examine predictive accuracy on four medical databases: Pima Indian diabetes, breast cancer, heart disease, and gall bladder removal. ARTMAP-IC results are equal to or better than those of logistic regression, K nearest neighbour (KNN), the ADAP preceptron, multisurface pattern separation, CLASSIT, instance-based (IBL), and C4. ARTMAP dynamics are fast, stable, and scalable. A voting strategy improves prediction by training the system several times on different orderings of an input set. Voting, instance counting, and distributed representations combine to form confidence estimates for competing predictions.

Fractal modeling and segmentation for the enhancement of microcalcifications in digital mammograms

The objective of this research is to model the mammographic parenchymal, ductal patterns and enhance the microcalcifications using deterministic fractal approach. According to the theory of deterministic fractal geometry, images can be modeled by deterministic fractal objects which are attractors of sets of two-dimensional (2-D) affine transformations. The iterated functions systems and the collage theorem are the mathematical foundations of fractal image modeling. In this paper, a methodology based on fractal image modeling is developed to analyze and model breast background structures. We show that general mammographic parenchymal and ductal patterns can be well modeled by a set of parameters of affine transformations. Therefore, microcalcifications can be enhanced by taking the difference between the original image and the modeled image. Our results are compared with those of the partial wavelet reconstruction and morphological operation approaches. The results demonstrate that the fractal modeling method is an effective way to enhance microcalcifications. It may also be able to improve the detection and classification of microcalcifications in a computer-aided diagnosis system.

Segmentation and numerical analysis of microcalcifications on mammograms using mathematical morphology

The top-hat and watershed algorithms of mathematical morphology have been applied to detect automatically and segment microcalcifications on mammograms digitized to a pixel resolution of 40 microns using a CCD camera. The database comprised 38 cases from the breast assessment clinic in Liverpool. For all cases, both craniocaudal (CC) and lateral oblique (LO) views were available. 19 cases were proven to be benign and 19 malignant based on cytology and histology. Malignant clusters contained more microcalcifications (14 malignant, 10 benign), occupied a larger area (37 mm2, 9 mm2) and had longer cluster perimeters than benign clusters (33.2 mm, 15.5 mm). Malignant microcalcifications exhibited a wider variety of shapes and were more heterogeneous in terms of image signal intensity than benign microcalcifications. Further mathematical morphology algorithms were applied to describe microcalcification shape in terms of the presence or absence of infoldings, elongation, narrow irregularities and wide irregularities. The three largest microcalcifications were selected for each case and, using a "leave-one-out" approach, each microcalcification was classified in respect of its five nearest neighbours as either malignant or benign. The area under the curve of a receiver operating characteristic (ROC) analysis of the proportion of the three microcalcifications which agreed with the true diagnosis increased from 0.73 (CC) and 0.63 (LO) to 0.79 when both views were considered. Next, each cluster in turn was ranked according to its agreement with the database as a whole over 21 features. An ROC analysis was performed to investigate the effect on sensitivity and specificity of the proportion of the nine nearest neighbours that agreed with the true classification. The largest area under the ROC curve was 0.84 produced by the four features of proportion of irregular microcalcifications, proportion of round microcalcifications, number of microcalcifications in the cluster and the interquartile range of microcalcification area. The shape of microcalcifications is confirmed as being of overriding importance in classifying cases as either malignant or benign. This observation motivates a further study enhanced by using magnified views digitized to a higher resolution by a laser scanner. This will enable the reliable assessment of the shape of a greater number of microcalcifications in each cluster, which is likely to increase further the discriminating power of the image analysis routines and lead to the development of an expert system for automatic mammographic screening.

An improved computer-assisted diagnostic scheme using wavelet transform for detecting clustered microcalcifications in digital mammograms

RATIONALE AND OBJECTIVES

We evaluated the potential usefulness of a computer-assisted diagnostic (CAD) scheme incorporating the wavelet transform for detecting clustered microcalcifications in mammograms.

METHODS

A wavelet transform technique was applied to the detection of clustered microcalcifications. We examined several wavelets to study their effectiveness in detecting subtle microcalcifications. We used a database consisting of 39 mammograms containing 41 clusters of microcalcifications. The performance of the wavelet-based CAD scheme was evaluated using free-response receiver operating characteristic analysis.

RESULTS

The CAD scheme with the wavelet transform was useful in detecting some of the subtle microcalcifications that were not detected by our previous scheme, which was based on the difference-image technique. When the two schemes were combined, the overall performance was improved to a sensitivity of approximately 95%, with a false-positive rate of 1.5 clusters per image.

CONCLUSION

The wavelet transform approach can improve the detection of subtle clustered microcalcifications.

Reduction of false positives in lung nodule detection using a two-level neural classification

The authors have developed a neural-digital computer-aided diagnosis system, based on a parameterized two-level convolution neural network (CNN) architecture and on a special multilabel output encoding procedure. The developed architecture was trained, tested, and evaluated specifically on the problem of diagnosis of lung cancer nodules found on digitized chest radiographs. The system performs automatic "suspect" localization, feature extraction, and diagnosis of a particular pattern-class aimed at a high degree of "true-positive fraction" detection and low "false-positive fraction" detection. In this paper, the authors aim at the presentation of the two-level neural classification method in reducing false-positives in their system. They employed receiver operating characteristics (ROC) method with the area under the ROC curve (A(z)) as the performance index to evaluate all the simulation results. The two-level CNN showed superior performance (A(z)=0.93) to the single-level CNN (A(z)=0.85). The proposed two-level CNN architecture is proven to be promising and to be extensible, problem-independent, and therefore, applicable to other medical or difficult diagnostic tasks in two-dimensional (2-D) image environments.

Artificial convolution neural network techniques and applications for lung nodule detection

We have developed a double-matching method and an artificial visual neural network technique for lung nodule detection. This neural network technique is generally applicable to the recognition of medical image pattern in gray scale imaging. The structure of the artificial neural net is a simplified network structure of human vision. The fundamental operation of the artificial neural network is local two-dimensional convolution rather than full connection with weighted multiplication. Weighting coefficients of the convolution kernels are formed by the neural network through backpropagated training. In addition, we modeled radiologists' reading procedures in order to instruct the artificial neural network to recognize the image patterns predefined and those of interest to experts in radiology. We have tested this method for lung nodule detection. The performance studies have shown the potential use of this technique in a clinical setting. This program first performed an initial nodule search with high sensitivity in detecting round objects using a sphere template double-matching technique. The artificial convolution neural network acted as a final classifier to determine whether the suspected image block contains a lung nodule. The total processing time for the automatic detection of lung nodules using both prescan and convolution neural network evaluation was about 15 seconds in a DEC Alpha workstation.