Diagnostic Performance of AI for Recurrent Cancers in Breast Conserved Patients: A Retrospective Analysis

Abstract

itle: Diagnostic Performance of AI forRecurrentCancers inBreast Conserved Patients: A Retrospective AnalysisIntroduction and Aim Advances in artificial intelligence (AI)technologies have led to techniques that could increase breast cancer detection, improve clinical efficiency in breast imaging practices, and guide decision-making regarding screening and prevention strategies(1). Last year Kim et al. developed and validated an AI algorithm by using large-scale data and showed better diagnostic efficiency than radiologists in breast cancer detection (2).Our aim was to evaluate the performance of this (2) AIalgorithm in detecting recurrences and post-operative changes in post-operative screening.Materials and MethodsIn total, 253postoperative mammograms were extracted from thearchive system of our oncology institute.Two hundred forty-two of them were diagnosed asnon-recurrent post-operative mammograms with a follow-up for at least two years, elevenof them were biopsy-proven recurrentmammograms. Mammograms were reported by a six years experienced breast radiologist. We used a recently developed diagnostic support software (Lunit INSIGHT MMG, Seoul, South Korea) on a free website (https://insight.lunit.io/mmg/login). After uploading DICOM images of four positions bilateral mammograms, the system calculates an abnormality score(0-15:low, 16-50:moderate, 51-100:high)which reflects the likelihood of malignancy of the detected lesion. Low and moderate results were accepted as negative, and high results were accepted as positive mammograms. The sensitivity and specificity values of the AI algorithmwere calculated.ResultsAI algorithm correctly detected 9 of 11 recurrent mammograms and 202 of 242 normal post-operative mammograms. The number of false-positive mammograms were 40, and false-negative mammograms were 2. Sensitivity and specificity values of AI in detecting recurrences were 81%, 83% respectively. According to a detailed analysis of false-positive results; 2 of 40 false-positive results were detected in the non-operated contralateral breast which was reported as BIRADS-2 by radiologists. Thirty-eightfalse positive results which were detected in the operated breast consisted of 36 post-operative scar tissue,1 focal asymmetry, and 1 simple cyst. According to a detailed analysis of false-negative results; both of them were non-calcified new lesions in non-operated contralateral breasts. Discussion and ConclusionIn conclusion although, presence of a need for prospective clinical trials, AI can not potentially enhance the capacity of post-operativescreeningbyincreasing false positiverates. AI systems require thorough evaluation in the clinical setting with history of patientsbefore widespread adoption.KeywordsArtificial Intelligence,References1-Bahl M. Artificial Intelligence: A Primer for Breast Imaging Radiologists. J Breast Imaging. 2020 Aug;2(4):304-314. doi: 10.1093/jbi/wbaa033. 2-Kim HE, Kim HH, Han BK, et al. (2020) Changes in cancer detection and false-positive recall in mammography using artificial intelligence: a retrospective, multireader study. Lancet Digit Health. 2: e138-e148.