Concept: Automated asymmetry detection in digital mammography using u-net deep learning network

Introduced by Ensi Khalili Pooya, n30.khalili@gmail.com and Ashkan Yousefi, PhD, ashkan.yousefi@berkeley.edu.

Introduction:

Population-based breast cancer screening programs with mammography have proven to reduce mortality and morbidity associated with advanced stage of disease.

Screening mammography included two views of each breast (craniocaudal-CC and mediolateral oblique-MLO). One of the suspicious mammographic findings for mass is asymmetries, findings that represent unilateral deposition of fibroglandulair tissue not confirming completely to definition of a mass. Asymmetries are classified in four groups: 1. Asymmetry : as an area of fibroglandulair tissue visible on only one mammographic projection, mostly caused by superimposition of normal breast tissue.2. Focal Asymmetry: visible on two projections, hence a real find in rather than superposition.3. Global Asymmetry: consisting of an asymmetry over at least one-quarter of breast and is usually a normal variant.4. Developing Asymmetry: new, larger and more conspicuous than on a previous examination.

Among these four types we want to work on type 2(Focal Asymmetry) because this type has to be differentiated from the mass. For this reason radiologists request additional view, Focal Compression Magnification view (FCMV), from the desired Focal Asymmetry to see if there is a real mass under it(asymmetry does not resolve in FCMV) or not . Each additional view has radiation dose equal to one chest X-ray.

In this study we aim to produce AI algorithm to detect Focal Asymmetries, with accuracy similar to an expert radiologist, which can predict the probability of being resolved in FCMV (low/high probability). This computer-aided detection system help radiologist to evaluate mammographic exams faster and also reduce the number of requests for FCMV in order to reduce radiation exposure dose to patients.

Patient selection :

This study will conduct with anonymized data retrospectively collected from our institutional archive.

Digital mammogram exams from women attending the screening program at our hospital for diagnostic purposes from 2016 to present.

All exams are bilateral and included two views (craniocaudal CC and mediolateral oblique MLO).

The images were acquired by HOLOGIC, USA.

Conclusion:

In this abstract we presented computer-aided detection and decision system which can detect Focal Asymmetries in digital mammography and can predict which one has low probability to resolve in FCMV so radiologists must take additional projection from them and which one has high probability and radiologists can overlook.

This strategy reduces exposure radiation dose to patients and also reduce the number of patients recall to a hospital and reduce costs.