Vision-Model-Based Image Enhancement for Digital Mammography

摘要

The objective was to evaluate the utility of a visual discrimination model, JNDmetrix, for predicting/optimizing the effects of grayscale window and level (WIL) parameters on the detectability of breast lesions. Two observer performance studies were conducted to correlate detection performance with JNDmetrix predictions of lesion conspicuity. In the first set of experiments, 2AFC detection thresholds were measured with five WIL conditions applied to simulated mammographic backgrounds and 'lesions' using nonmedical observers. The detectability of real microcalcification clusters in digitized mammograms was then evaluated for three WIL conditions in an ROC study with mammographers. For the simulated lesions/backgrounds, the model correctly predicted WIL conditions that minimized the detection thresholds, supporting our hypothesis that the model could be used to automate the selection of optimal W/L settings. Experimental and model results showed significant reductions in thresholds when WIL processing was applied locally near the lesion. ROC results with digitized mammograms read by radiologists failed to show enhanced detection of microcalcifications using a localized W/L frame. This was probably due to the nonuniform appearance of parenchymal tissue across the image. Future experiments will provide readers both the uniform contextual information of the full image and enhanced lesion contrast within a localized W/L frame.