Guiding search in 3D volumes with 2D synthesized images

Abstract

Introduction: Three-dimensional volumetric images have become prevalent for cancer detection tasks by allowing radiologists to segment features of interest from anatomical background noise (e.g., breast parenchyma). However, spatially small 3D targets that are hard to detect in the visual periphery are often not foveated and then missed. Radiologists are routinely presented with a 2D synthesized image (2D-S) of the 3D volumetric images to assist the 3D search. We investigate the visual-cognitive processes by which the 2D-S benefits the 3D search for small and large targets. Methods: Six trained observers participated in a hybrid Yes/No, rating, and localization search task. They searched for either a small (0.15 deg) or large (0.5 deg) target in 3D volumes (100 slices) with or without an additional 2D-S. We applied a high pass filter and a pixel-wise maximum operation across the 3D slices to produce the 2D-S image. Results: The presence of the 2D-S image improved observers’ hit rate (ΔHR= 0.31, p<.001) and correct localization rate (ΔCL= 0.278, p<.001) of small targets in the 3D images. For larger targets, there was a smaller improvement (ΔHR= 0.115, p>.05). Eye movement analysis showed significantly fewer search errors in the 2D-S + 3D condition relative to the 3D condition for the small target (M3D+2D-S= .15, M3D= .45, p<.001). The presence of the 2D-S image resulted in a pattern of search consistent with guided drilling behavior: fewer eye movements to foveate the small target (M3D+2D-S = 24.5, M3D = 44.9, p<.05), fewer reversals in scrolls (M3D+2D-S = 7.5, M3D = 14.29, p<.01), and shorter response times (M3D+2D-S = 22.81 sec., M3D = 34.92 sec., p<.01). Conclusion: Our results suggest that 2D synthetic images effectively guide 3D search towards likely target locations, resulting in guided drilling behavior and reduced search errors for small targets difficult to detect in the periphery.