Technical feasibility of automated blur detection in digital mammography using convolutional neural network

• DOI: 10.1186/s41747-024-00527-0
• Publication Date: 2024-11-18T15:56:47Z
• Authors: S. Nowakowska; V. Vescoli; T. Schnitzler; C. Ruppert; K. Borkowski; A. Boss; C. Rossi; B. Wein; A. Ciritsis
• Keywords: Ground truth; Gaussian blur; Digital Mammography

Abstract Background The presence of a blurred area, depending on its localization, in mammogram can limit diagnostic accuracy. goal this study was to develop model for automatic detection blur diagnostically relevant locations digital mammography. Methods A retrospective dataset consisting 152 examinations acquired with mammography machines from three different vendors utilized. areas were contoured by expert breast radiologists. Normalized Wiener spectra (nWS) extracted sliding window manner each mammogram. These served as input convolutional neural network (CNN) generating the probability originating region. resulting mask, upon thresholding, facilitated classification either or sharp. Ground truth test set defined consensus two Results significant correlation between view ( p < 0.001), well laterality and = 0.004) identified. developed AUROC 0.808 (95% confidence interval 0.794–0.821) aligned 78% (67–83%) mammograms classified blurred. For sharp, achieved agreement 75% them. Conclusion assessed. results indicate that robust approach detection, based feature extraction frequency space, tailored radiologist expertise regarding clinical relevance, could eliminate subjectivity associated visual assessment. Relevance statement This model, if implemented practice, provide instantaneous feedback technicians, allowing prompt retakes ensuring only high-quality are sent screening tasks. Key Points Blurring limits interpretation objective tool ensures image quality, reduces unnecessary exposures. spectrum analysis CNN enabled automated Graphical