Malignancy Detection on Mammography Using Dual Deep Convolutional Neural Networks and Genetically Discovered False Color Input Enhancement

Abstract

Breast cancer is the most prevalent malignancy in the US and the third highest cause of cancer-related mortality worldwide. Regular mammography screening has been attributed with doubling the rate of early cancer detection over the past three decades, yet estimates of mammographic accuracy in the hands of experienced radiologists remain suboptimal with sensitivity ranging from 62 to 87% and specificity from 75 to 91%. Advances in machine learning (ML) in recent years have demonstrated capabilities of image analysis which often surpass those of human observers. Here we present two novel techniques to address inherent challenges in the application of ML to the domain of mammography. We describe the use of genetic search of image enhancement methods, leading us to the use of a novel form of false color enhancement through contrast limited adaptive histogram equalization (CLAHE), as a method to optimize mammographic feature representation. We also utilize dual deep convolutional neural networks at different scales, for classification of full mammogram images and derivative patches combined with a random forest gating network as a novel architectural solution capable of discerning malignancy with a specificity of 0.91 and a specificity of 0.80. To our knowledge, this represents the first automatic stand-alone mammography malignancy detection algorithm with sensitivity and specificity performance similar to that of expert radiologists.

Keywords: Convolutional neural networks; Deep learning; Machine learning; Mammography.

Fig. 1

In generational order from left to right, the same image of a red kite is transformed by the current generation of image preprocessing, created by the evolutionary system described. Images courtesy of PT and Prof. Abigail Morrison

Scheme 1

Comparing accuracy of unenhanced and enhanced input training. Accuracy score against training time for region of interest binary classification of malignancy risk (p < 0.005)

Fig. 2

Enhanced inputs: normal window patch (a), window patch containing malignancy (b), normal full image (c), and full image containing malignancy and scarring (d). All images shown are from different patients

Fig. 3

Varying the CLAHE window and clipping parameters differently across the color channels affords useful enhancement across a wider range of fidelity resolutions, across the majority of the breast. Affording sharper structural enhancement over more scales of resolution and windows of tonal range

Fig. 4

Illustration showing the dimensions of the sliding window region, traversed across the full image, used to generate the map of regions of interest, and eventually their local probability of risk

Fig. 5

System architecture demonstrating full mammogram and patch input into dual deep CNN instances with an additional final random forest analytic component

Scheme 2

Final classification receiver operator characteristic curve