RAI Labs Duke

Computer-aided detection: Breast imaging

Approximately 20% of all breast cancers are missed in mammography screening. Double reading by 2 radiologists can improve sensitivity of cancer detection, but that practice is not feasible in the US. As an alternative, computer-aided detection (CAD) has been developed as an automated second reader. Commercial CAD systems have gained widespread acceptance in recent years, with up to 20% improvement in cancer detection and over 90% sensitivity for clustered microcalcifications. Currently CAD products still detect only about 80% of breast masses, which account for half of all breast cancers. This research seeks to address this clinically significant challenge.

We have designed several computer aided detection (CAD) systems to help radiologists with early detection of subtle masses. Each CAD system consists of two stages. The first stage uses an image processing filter to detect suspicious regions with high sensitivity and low specificity (steps A and B in the flowchart here). The second stage extracts digital image features based on morphology, texture, and vision-based Hotelling observers (steps C, D, and E in the flowchart). Linear discriminant statistical models merged these features together to predict whether each suspicious region is an actual breast mass or not. The final system performs with up to 90% sensitivity over 1400+ DDSM mammograms. Incorporating this type of a CAD system into mammographic screening may increase the early detection of breast masses.

We relied extensively on DOD’s Digital Database for Screening Mammography (DDSM) breast cancer database for this study, as it is the largest publicly available database of its kind, providing digitized mammograms with proven outcomes. We used 1406 images including 215 containing 238 benign masses, 234 containing 248 malignant masses, and 957 normal cases.

We identified approximately 9000 suspicious regions using a difference of gaussians image processing filter. For each region, we extracted 36 commonly used digital image features describing its morphology and texture. In addition, we developed several novel vision-based Hotelling observers (HO). The HO is the optimal linear detector for a known signal given information about the signal, background, and covariance matrix. Hotelling observers effectively track detection performance of human observers such as radiologists. These observers included the sub-region Hotelling observer and Laguerre-Gauss channelized Hotelling observer (LG-CHO).

Currently, our research has yielded the following conclusions:

  • Investigated predictive models to detect suspicious breast masses in digitized mammograms.
  • Vision-based Hotelling observer models improve performance over traditional models based on morphology and texture features.
  • Use of truly independent validation is needed over traditional cross-validation (train and test) sampling.
  • Size of DDSM database was large enough to allow independent validation, resulting in robust models which generalize well to new cases.
  • Models deliver up to 90% sensitivity but with lower specificity than commercial systems, which is an area of further research.

Example Images:

77 year old woman with ill-defined, irregular mass (left shows mammogram with radiologist's hand-drawn outline of lesion). CAD detects the mass as well as generating 2 false positives in this image (right).

48 year old woman with ill-defined/spiculated, irregular mass (left). CAD placed 2 true positive marks in the lesion and generated 2 other false positive marks (right).

38 year old woman with ill-defined, irregular mass (left). CAD placed 2 true positive marks in the lesion and generated no false positive marks (right).