A perceptual study in DCE-MRI for breast cancer diagnosis: First phase in the development of a clinically meaningful decision support system
These file attachments have been under embargo and were made available to the public after the embargo was lifted on 5 August 2011.
Rational: An experiment is described whose purpose is to extract those image characteristics that radiologists use in diagnostic decision making. It is one phase in the development of a clinical decisionsupport system to assist radiologists using DCE-MRI (dynamic contrast-enhanced MRI) for breast cancer diagnosis. The results of this experiment will be used to develop a case-based reasoning system, whichrelies on presenting prior similar cases with known diagnosis froma database to aid decision making. Methods: Clinical similarity formass lesions was established by four expert radiologists who systematically sorted lesions visualized by DCE-MRI into similarity clusters using a proprietary software tool. Cognitive analysis was used toidentify the relevant perceptual features characterizing each cluster, such that the list of features and clusters define the clinicalsimilarity. There were no constraints on the number or size of clusters that could be created. The radiologists first individually clustered a total of 214 lesions. A subsequent phase required all radiologists to agree on both a cluster designation and assignment of eachlesion into a cluster. Results: Radiologists created individually10, 10, 12, and 16 clusters. Of this initial cluster assignment, there was unanimous agreement in ~20% of the lesions, and majority agreement of ~60%. The final consensus assignment created 16 clusters; two consisted of all malignant lesions; two consisted of a majority of benign lesions, three were large approximately equal mixes of benign and malignant lesions, while the remaining nine were small clusters representing lesions with clinically relevant special characteristics of low clinical prevalence. Conclusions: The cognitive analysis revealed that the image characteristics differentiating the clusters are highly correlated to BI-RADS lesion descriptors. The radiologists were excellent in clustering certain malignant lesions, very good with some benign lesions; while as expected there was large variability in the majority of lesions.