Binghamton University

Computer Science Department

Dr. Barbara Croft

Diagnostic Imaging Program

National Cancer Institute

Breast Cancer is second only to lung cancer as a tumor-related cause of death in women. More that 180,000 new cases are reported annually in the US alone and, of these, 43,900 women and 290 men died last year. Furthermore, the American Cancer Society estimates that at least 25% of these deaths could be prevented if all women in the appropriate age groups were regularly screened.

Although there exists reasonable agreement on the criteria for benign/malignant diagnoses using fine needle aspirate (FNA) and mammogram data, the application of these criteria are often quite subjective. Additionally, proper evaluation of FNA and mammogram sensor data is a time consuming task for the physician. Inter-and-inter-observer disagreement and/or inconsistencies in the FNA and mammogram interpretation further exacerbate the problem.

Consequently, Computer Aided Diagnostics (CAD) in the form of Evolutionary Programming derived neural networks, neural network hybrids and neural networks operationg alone (utilized for pattern recognition and classification) offer significant potential to provide an accurate and early automated diagnostic technology. This automated technology may well be useful in further assisting with other problems resulting from physical fatigue, poor mammogram image quality, inconsistent FNA discriminator numerical assignments, as well as other possible sensor interpretation problems.

The purpose of this proposed session is to present current and ongoing research in the CAD of breast carcinoma. Specifically, this session has the following objectives:

• To show the advantages of using Evolutionary Programming and Neural Networks as an aid in the breast cancer diagnostic process
• To demonstrate the application of EP and neural network hybrid systems in solving the Inter-Observability problem.
• To establish that CAD tools are simple and economical to implement in the clinical setting
• To demonstrate that CAD tools can provide the cytophathogists, radiologists and neurosurgeons with an early diagnosis of breast cancer that is accurate, consistent and efficient as well as accessible.

Some practical results of CAD of breast cancer sensor data using neural networks are expected to be:

• Operational software which will aid the physician in making the diagnosis, quite possibly in real time, and once formulated and tested, they are always consistent, not prone to human fatigue or bias.
• Providing diagnostic assistance for the intra-and-inter-observability problems by ultimately minimizing the subjective component of the diagnostic process
• Providing an initial detection and/or classification process in the absence of a qualified physician
• Providing possible (and probably currently unknown) relationships between sensor environment discriminators and a correct diagnosis.

This session is comprised of the following five invited papers:

1. A Status Report on Identifying Important Features for Mammogram Classification

D.B. Fogel, (Natural Selection, Inc.)

E.C. Wasson (Maui Memorial Hospital)

E.M. Boughton, (Hawaii Industrial Laboratories)

V.W. Porto and P.J. Angeline (Natural Selection, Inc.)

Disagreement or inconsistencies in mammographic interpretation motivates utilizing computerized pattern recognition algorithms to aid the assessment of radiographic features. We have studied the potential for using artificial neural networks (ANNs) to analyze interpreted radiographic features from film screen mammograms. Attention was given to 216 cases (mammogram series) that presented suspicious characteristics. The domain expert (Wasson) quantified up to 12 radiographic features for each case based on guidelines from previous literature. Patient age was also included. The existence or absence of malignancy was confirmed in each case via open surgical biopsy. The ANNs were trained using evolutionary algorithms in a leave-one-out cross validation procedure. Results indicate the ability for small linear models to also provide reasonable discrimination. Sensitivity analysis also indicates the potential for understanding the networks’ response to various input features.

2. Application of Artificial Neural Networks for Diagnosis of Breast Cancer

J.Y. Lo and C.E. Floyd (Digital Imaging Research Division, Dept. of Radiology, Duke Univ. Medical Center, and Dept. of Biomedical Engineering, Duke Univ.)

We will present several current projects pertaining to artificial neural networks (ANN) computer models that merge radiologist-extracted findings to perform computer-aided diagnostics (CADx) of breast cancer. These projects include (1) the prediction of breast lesion malignancy using mammographic and patient history findings, (2) the further classification of malignant lesions as in situ carcinoma vs. invasive cancer, (3) the prediction of breast cancer utilizing ultrasound findings, and (4) the customization and evaluation of CADx models in a multi-institution study. Methods: These projects share in common the use of feedforward, error backpropagation ANNs. Inputs to the ANNs are medical findings such as mammographic or ultrasound lesion descriptors and patient history data. The output to the ANN is the biopsy outcome (benign vs. malignant, or in situ vs. invasive cancer) which is being predicted. All ANNs undergo supervised training using actual patient data. Performance is evaluated by ROC area, specificity for a given high sensitivity, and/or positive predictive value (PPV). Results: We have developed ANNs, which have the ability to predict the outcome of breast biopsy at a level comparable or better than expert radiologists. For example, using only 10 mammographic findings and patient age, the ANN predicted malignancy with a ROC area of 0.86 = B1 0.02, a specificity of 42% at a given sensitivity of 98%, and a 43% PPV. onclusion: These ANN decision models may assist in the management of patients with breast lesions. By providing information which was previously available only through biopsy, these ANNs may help to reduce the number of unnecessary surgical procedures and their associated cost. Contributions made by this abstract: This abstract describes the application of simple backprop ANNs to a wide range of predictive modeling tasks in the diagnosis of breast cancer. The group is one of the most authoritative in the field of computer-aided diagnosis, with a tract record that encompasses many radiological imaging modalities and engineering disciplines.

3. Optimizing the Effective Number of Parameters in Neural Network Ensembles: Application to Breast Cancer Diagnosis

Y. Lin and X. Yao, ( Dept of Computer Science, Australian Defense Force Academy, Canberra)

The idea of negative correlation learning is to encourage different individual networks in an ensemble to learn different parts or aspects of the training data so that the ensemble can learn the whole training data better. This paper develops a technique of optimizing the effective number of parameters in a neural network ensemble by negative correlation learning. The technique has been applied to the problem of breast cancer diagnosis.

4. Artificial Neural Networks in Breast Cancer Diagnosis: Merging of Computer-Extracted Features from Breast Images

Maryellen L. Giger

Kurt Rossmann Laboratories for Radiologic Image Research

The University of Chicago

Department of Radiology MC2026

5841 S. Maryland Ave.

Chicago, IL 60637

m-giger@uchicago.edu

Phone: 773-702-6778

Fax: 773-702-0371

Computer-aided diagnosis (CAD) can be defined as a diagnosis made by a radiologist who takes into consideration the results of a computerized analysis of radiographic images and uses them as a "second opinion" in detecting lesions, assessing disease, and/or in making diagnostic decisions. The final decision regarding diagnosis and patient management would be made by the radiologist. CAD is analogous to a spell checker. In mammography, computerized methods for the detection and characterization of lesions are being developed to aid in the early detection of breast cancer. Since mammography is becoming a high volume x-ray procedure routinely interpreted by radiologists and since radiologists do not detect all cancers that are visible on images in retrospect, it is expected that the efficiency and effectiveness of screening could be increased by CAD. In addition, computerized methods are being developed to aid in the characterization of lesions in order to potentially reduce the number of benign cases sent to biopsy.

Once a lesion is detected, the radiologist's decision concerns patient management -- that is, return to routine screening, perform follow-up mammograms, or send to biopsy? This differs from a purely benign versus malignant determination. Many breast cancers are detected and referred for surgical biopsy on the basis of a radiographically detected mass lesion or cluster of microcalcifications. Although general rules for the differentiation between benign and malignant mammographically identified breast lesions exist, considerable misclassification of lesions can occur. On average, less than 30% of masses referred for surgical breast biopsy are actually malignant. We have been developing computerized analysis schemes to aid in distinguishing between malignant and benign lesions. Such methods can use features extracted either by computer or by radiologists. These features are then merged by classifiers, such as linear discriminant functions or artificial neural networks, to yield estimate of the likelihood of malignancy. These computerized methods for the characterization of breast lesions are being developed for mammography, sonography, and magnetic resonance imaging.

5. Investigation of and Preliminary Results for the Solution of the Inter-Observability Problem using Fine Needle Aspirate (FNA) Data

W. H. Land, JR and L. Loren (Computer Science Dept., Binghamton Univ. and T. Masters (TMAIC, Vestal, NY)

This paper provides a preliminary evaluation of the accuracy of Computer Aided Diagnostics (CAD) in addressing the inconsistencies of Inter-Observability scoring. The Inter-Observability problem generally relates to different cytopathologists and radiologists, etc. at seperate locations scoring the same type of samples differently using the same methodologies and environment discriminates. Two different approaches are currently being investigated: (1) a recently developed Evolutionary Programming (EP) / Probabilistic Neural Network (PNN) hybrid, and (2) a classification model based on the thresholding of means of all predictors called the "mean of predictors" model. Method: Two distinctly different FNA data sets were used. The first was the data set collected at the Univ. of Wisconsin (Wolberg data set) while the other was a completely independent one defined and processed at the Breast Care Center, Syracuse University (Syracuse dataset). Results of several experiments performed using the EP/PNN hybrid (which provided several unique network configurations) are first summarized. The EP/PNN hybrid was trained on the Wolberg dataset and the resultant models evaluated on the Syracuse dataset. For comparative purposes, these same hybrid architectures which were trained on the Wolberg set were also evaluated on the Wolberg validation set. The "mean of predictors" method first trained the thresholds using the original Wolberg training set. This model was then tested on the Wolberg test and validation sets, and on the Syracuse set. All three Wolberg datasets (train, test, validate) were then used to train the threshold, and this model was applied to the Syracuse data Results: Initial results using the EP/PNN hybrid showed a 85.2% correct classification accuracy with a 2.6% Type II (classifying malignant as benign) error averaged over five experiments when trained on the Wolberg data set and validated on the Syracuse data set. Training and validating on the Wolberg data set resulted in a 97% correct classification accuracy and a < 0.2% Type II error. These results are preliminary in that no attempt has been made to optimize the threshold setting. The paper will include several additional EP/PNN hybrid experimental results as well as optimum threshold settings and an ROC analysis. The "mean of predictors method" analysis produced the following preliminary results. Training the thresholds on the first 349 Wolberg samples resulted in a CAD model which provided: (1) a 98.8% classification accuracy and a 0% Type II error, (2) a 96% classification accuracy with a 1.7% Type II error when using the Wolberg test and validation sets respectively which confirms the EP/PNN preliminary results. Using the Syracuse validation set yielded a 96% classification accuracy and a 1% Type II error which is improved performance when compared with the EP/PNN results. Training the "mean of predictors" model on all 699 Wolberg samples and validating on the Syracuse dataset resulted in a 86% classification accuracy and a 1% Type II error. Again, these results match well with the EP/PNN results. We again emphasize these results are preliminary but very promising. Conclusions: Preliminary results using both the newly developed EP/PNN hybrid and the "mean of predictors" methods are very encouraging. We believe that both of these CAD tools will, with additional research and development effort, be useful additions to our growing group of CAD tools being developed at Binghamton University.