Abstract
PurposeThis work aimed to improve breast screening program accuracy using automated classification. The goal was to determine if whole image features represented in the discrete cosine transform would provide a basis for classification. Priority was placed on avoiding false negative findings.MethodsOnline datasets were used for this work. No informed consent was required. Programs were developed in Mathematica and, where necessary to improve computational performance ported to C++. The use of a discrete cosine transform to separate normal from cancerous breast tissue was tested. Features (moments of the mean) were calculated in square sections of the transform centered on the origin. K-nearest neighbor and naive Bayesian classifiers were tested.ResultsForty-one features were generated and tested singly, and in combination of two or three. Using a k-nearest neighbor classifier, sensitivities as high as 98% with a specificity of 66% were achieved. With a naive Bayesian classifier, sensitivities as high as 100% were achieved with a specificity of 64%.ConclusionWhole image classification based on discrete cosine transform (DCT) features was effectively implemented with a high level of sensitivity and specificity achieved. The high sensitivity attained using the DCT generated feature set implied that these classifiers could be used in series with other methods to increase specificity. Using a classifier with near 100% sensitivity, such as the one developed in this project, before applying a second classifier could only boost the accuracy of that classifier.
Highlights
Quality control systems have greatly improved the consistency of mammograms and technical advances have shortened exam time without negatively impacting performance [1]
Higher specificity and positive predictive value has been shown to correlate with more experience [3,4,5]
The abnormal findings rate was 8% and this was associated with a positive predictive value of 31.4%
Summary
Quality control systems have greatly improved the consistency of mammograms and technical advances have shortened exam time without negatively impacting performance [1]. The USA screening programs apparently operate at a sensitivity of 84.1% and a specificity of 90.4% [2]. Radiologist performance is an important component of program performance and no matter how skilled, reporting physicians will miss some cancers that, in retrospect, were visible in the mammogram [3]. One study reported a mean sensitivity of 77% with a range of 29% to 97%. Specificity ranged from 71% to 99% with an average of 90% [3]. Higher specificity and positive predictive value has been shown to correlate with more experience [3,4,5]. Sickles and colleagues [10] reported performance benchmarks based on an analysis of six breast cancer screening registries and more than 600 radiologists. The abnormal findings rate was 8% and this was associated with a positive predictive value of 31.4%
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have