Automatic Detection of Post-Operative Clips in Mammography Using a U-Net Convolutional Neural Network

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Surgical clips after breast conserving surgery (BCS) mark the tumor bed, indicating potential relapse areas.
• Accurate identification of these clips is crucial for effective follow-up imaging and patient management.

Purpose of the Study:

• To evaluate the efficacy of a U-Net-based deep convolutional neural network (dCNN) in detecting surgical clips on follow-up mammograms post-BCS.
• To compare the AI's performance against human readers in identifying these surgical markers.

Main Methods:

• A U-Net dCNN was trained and validated on a dataset of mammograms and tomosynthesis images containing surgical clips and calcifications.
• An external test set was annotated by experienced radiologists, and the AI's performance was assessed using accuracy and interrater agreement (Cohen's Kappa).

Main Results:

• The AI model achieved classification accuracy between 88.2% and 92.6% on the validation set, comparable to human readers.
• The AI demonstrated substantial agreement with radiologists (Cohen's Kappa: 0.72-0.78), while radiologists showed near-perfect agreement (0.84).
• A misclassification rate of 17.4% was observed for calcifications mistaken as clips.

Conclusions:

• AI techniques, specifically U-Net dCNNs, can effectively identify surgical clips in mammograms following BCS.
• This AI application holds potential for improving patient triage and automating quality management workflows, such as PGMI evaluation.