Accession Number : ADA302530

Title :   Computer-Aided Mammography Using Automated Feature Extraction for the Detection and Diagnosis of Breast Cancer.

Descriptive Note : Annual rept. 15 Sep 94-14 Sep 95,

Corporate Author : DUKE UNIV MEDICAL CENTER DURHAM NC

Personal Author(s) : Lo, Joseph Y.

PDF Url : ADA302530

Report Date : 12 OCT 1995

Pagination or Media Count : 21

Abstract : We developed artificial neural network (ANN) techniques to predict breast lesion malignancy based on mammographic features extracted by radiologists. The 3-layer backpropagation ANNs were trained and tested using the round robin technique and evaluated by ROC (receiver operating characteristic) analysis. Using all 11 available features from 206 patients, the ANN performed with ROC area AZ of 0.84 j 0.03, which was not significantly different from the expert radiologists' AZ of 0.85 + or - 0.03 (2-tailed p-value =0.54). We then ranked the importance of individual features to reduce the number of ANN input features. The resulting 6-feature ANN had AZ of 0.86 + or - 0.03 which was still not significantly different than that of the expert radiologists with p =0.34. The result was an optimally simplified ANN for merging features to predict breast lesion malignancy. In the following years, work will focus on automated extraction of those features to feed into the ANN inputs, thus producing a fully automated computer-aided diagnosis system.

Descriptors :   *NEURAL NETS, *DIAGNOSIS(MEDICINE), *COMPUTER AIDED DIAGNOSIS, *MAMMARY GLANDS, *BREAST CANCER, INPUT, AUTOMATION, DETECTION, EXTRACTION, FEEDING, LESIONS.

Subject Categories : Medicine and Medical Research
      Anatomy and Physiology
      Cybernetics

Distribution Statement : APPROVED FOR PUBLIC RELEASE