Use of a 'Mathematical Microscope' to Understand Radiologists' Errors in Breast Cancer Detection
Honorable Mention - Imaging InformaticsClaudia Mello-Thoms ; University of Pittsburgh;
Content:
Excluding cancers of the skin, breast cancer is the most common type of cancer for women in the US. Mammography screening is the only method proven to reduce mortality from this disease, but due to several factors, between 10-30% of breast cancers are still missed at screening. These misses have been divided into: i) search errors (lesions did not attract any amount of visual attention); ii) perceptual errors (lesions attracted visual attention but not long enough for object recognition); iii) decision-making errors (lesions attracted visual attention for a long period of time, but were ultimately dismissed). Perceptual and decision-making errors account for 70% of all misses. However, we need to determine how these lesions differed from correctly reported breast cancers.
Technology:
Using a mathematical microscope , a wavelet packets transform, we have derived a model that characterizes, like the human visual system, each area fixated. The basic steps of the decomposition are i) area segmentation into squares measuring 5° of visual field (size of the fovea); ii) wavelet packets filtering; iii) feature extraction; iv) normalization. Each miss and correctly reported cancer is characterized by this local representation and by a representation of the background areas used to support the decision.
Design:
Four MQSA-certified radiologists read 40 two-view digitized mammogram cases, of which 30 cases contained malignant masses. Eye-position and decision response were recorded. Cancers that were correctly reported were contrasted with those that attracted visual attention but were not reported.
Results:
Analysis of Variance of local representation of cancers that were correctly reported and those that yielded perceptual or decision-making errors do not yield statistically significant differences, although differences between correct reports and perceptual errors reach borderline significance for certain orientations/spatial frequency ranges (ex, Scheffes post-hoc test, p=0.054). Analysis of background areas used for decision support show no difference in search strategy in case of perceptual errors (indicating no object recognition) and statistically significant differences after fixating both correctly reported cancers and decision-making errors.
Conclusion:
Local feature analysis often cannot explain why lesions were missed, but a global analysis that takes into account support information can characterize why errors occur.
