Deep Learning or Fundamental Descriptors?

Abstract

HomeRadiologyVol. 287, No. 2 PreviousNext CommunicationsFree AccessLetters to the EditorDeep Learning or Fundamental Descriptors?Daniel B. Kopans Daniel B. Kopans Author AffiliationsDepartment of Radiology, Breast Imaging Division, Massachusetts General Hospital, Harvard Medical School, 15 Parkman St, Suite 219, Boston, MA 02114e-mail: [email protected]Daniel B. Kopans Published Online:Apr 18 2018https://doi.org/10.1148/radiol.2017173053MoreSectionsPDF ToolsImage ViewerAdd to favoritesCiteTrack CitationsPermissionsReprints ShareShare onFacebookTwitterLinked InEmail Editor:Having placed great hope on computer-aided detection (CAD) back in the 1990s only to be disappointed in its ultimate failure to add much in breast imaging, I am still hopeful, but have a healthy skepticism, about “deep learning,” which is simply CAD on steroids. The article by Dr Bahl and colleagues (1) in the March 2018 issue of Radiology is interesting, but by mixing too many “high-risk lesions” (HRLs) together in the evaluation, as well as not analyzing the reasons for the computer results, the real value of the analysis is in question. As Dr Bahl and colleagues point out, the categories of lesions lumped (sorry) together reduces the ability to understand the value of the test. The HRLs included are widely differing in their etiology and histology as well as their risk for developing breast cancer and “upgrade” with excision. For example, the risk of upgrade for atypical ductal hyperplasia is much higher than that for flat epithelial atypia or a benign papilloma. This is clear in a recent article by two of the same authors (probably including many of the same patient lesions) showing that the upgrade rate for flat epithelial atypia by itself was extremely low at less than 3% (2). I suspect that they grouped the various HRLs together so that they had enough cases to be able to “train” the computer algorithms and then test them.What is not provided is the fact that if a trained radiologist reviewed the same cases and was asked to guess the rate of upgrade for each lesion, she or he would have likely done as well as the computer for many of the lesions. For example, if you performed a core biopsy of segmentally distributed pleomorphic calcifications and the pathologic examination revealed “atypical ductal hyperplasia” you would be fairly certain that the upgrade would reveal at least ductal carcinoma in situ. If, on the other hand, you biopsied and removed five or six borderline round calcifications that produced “atypical ductal hyperplasia,” you would likely be correct if you bet that the surgical excision would not lead to an upgrade.It is worth investigating the use of more powerful computers, but we need to remain skeptical until the data show a clear advantage. I would urge that investigators not simply report results, but be required to analyze histopathologically distinct lesions, separately, and not group lesions together as in this analysis. I would also urge that they try to understand what the computer was adding to a trained radiologist. It may be that simply by reading “pleomorphic” and “segmental” a computer algorithm will call for excision, and “deep learning” is not as deep as we would hope!Disclosures of Conflicts of Interest: disclosed no relevant relationships.