A full list of my publications is available in my on-line CV (curriculum vitae), including downloadable PDFs. The following are just a few selected papers to illustrate my fields of research interest.
PURPOSE: An artificial neural network was developed to predict breast cancer invasion based on BI-RADS mammographic findings and age. For patients classified as having invasive breast cancer, excisional biopsy may be obviated by obtaining histologic confirmation via stereotaxic needle core biopsy, and the patients may then undergo a single-stage surgery for mastectomy and/or axillary dissection.
MATERIALS AND METHODS: 266 biopsied lesions were randomly selected (96 malignant, 170 benign). Based on 9 BI-RADS mammographic findings and patient age, a 3-layer backpropagation network was developed to predict whether the 96 malignant lesions were in situ or invasive.
RESULTS. The ANN predicted invasion among malignant lesions with Az of 0.91 ± 0.03. It correctly identified all 28 cases of in situ cancer (100% specificity) and 48 of 68 invasive cancers (71% sensitivity).
CONCLUSION. This study presents an artificial neural network which predicts invasion of breast cancers using mammographic features and age. Previously available only through biopsy, this knowledge may assist in surgical planning for patients with breast lesions, and may help reduce the cost and morbidity of unnecessary biopsies.
RATIONALE AND OBJECTIVES. An artificial neural network (ANN) approach was developed for computer-aided diagnosis of mammography, using an optimally minimized number of input features.
METHODS. A backpropagation ANN merged 9 input features (age plus 8 radiographic findings extracted by radiologists) to predict biopsy outcome as its output. The features were ranked and more important ones were selected to produce an optimal subset of features.
RESULTS. Given all 9 features, the ANN performed with receiver operator characteristic (ROC) area index Az of 0.95 ± 0.01. Given only the 4 most important features, the ANN performed with Az of 0.96 ± 0.01. Although not significantly better than the ANN with all nine features, the ANN with the four optimized features was significantly better than expert radiologists' Az of 0.90 ± 0.02 (p = 0.01). This 4-feature ANN had 95% sensitivity and 81% specificity. For cases with calcifications, the radiologists' performance dropped to Az of 0.85 ± 0.04, while a specialized 3-feature ANN performed significantly better with Az of 0.95 ± 0.02 (p = 0.02).
CONCLUSIONS. An artificial neural network was developed for computer-aided diagnosis in mammography. Given only four input features, the network predicted biopsy outcome significantly better than expert radiologists, who also had access to other radiographic and non-radiographic data. The reduced number of features would substantially decrease data entry effort and potentially improve the networkÕs general applicability.
An adaptive linear element (Adaline) was developed to estimate the two-dimensional scatter exposure distribution in digital portable chest radiographs (DPCXR). DPCXRs and quantitative scatter exposure measurements at 64 locations throughout the chest were acquired for ten radiographically normal patients. The Adaline is an artificial neural network which has only a single node and linear thresholding. The Adaline was trained using DPCXR-scatter measurement pairs from five patients. The spatially invariant network would take a portion of the image as its input and estimate the scatter content as output. The trained network was applied to the other five images, and errors were evaluated between estimated and measured scatter values. Performance was compared against a convolution scatter estimation algorithm. The network was evaluated as a function of network size, initial values, and duration of training. Network performance was evaluated qualitatively by the correlation of network weights to physical models, and quantitatively by training and evaluation errors. Using DPCXRs as input, the network learned to describe known scatter exposures accurately (7% error) and estimate scatter in new images (< 8% error) slightly better than convolution methods. Regardless of size and initial shape, all networks adapted into radial exponentials with magnitude of 0.75, perhaps implying an ideal point spread function and average scatter fraction, respectively. To implement scatter compensation, the two-dimensional scatter distribution estimated by the neural network is subtracted from the original DPCXR.
[ work | CV | publications | personal ]