Source-linked AI summary
Robust breast cancer detection in mammography and digital breast tomosynthesis using annotation-efficient deep learning approach
William Lotter, Abdul Rahman Diab, Bryan Haslam, Jiye G. Kim, Giorgia Grisot, Eric Wu, Kevin Wu, Jorge Onieva Onieva, Jerrold L. Boxerman, Meiyun Wang, Mack Bandler, Gopal Vijayaraghavan, A. Gregory Sorensen
TL;DR
The paper addresses breast cancer model evaluation across mammography settings, including a population with low screening rates. It develops models for 2D mammography and DBT, with the proposed model achieving significantly higher AUC than comparison models.
Problem
Evidence is needed on breast cancer model performance across mammography settings, including populations with low screening rates.
Method
The approach trains models for 2D mammography and extends them to DBT using weakly supervised training from optimized images.
Results
The proposed model had a significantly higher AUC than the other models under comparison.
Takeaways & Limitations
The results support the proposed model as a promising approach for mammography and DBT evaluation across different screening contexts.
Abstract
from arXiv · showhide
Breast cancer remains a global challenge, causing over 1 million deaths globally in 2018. To achieve earlier breast cancer detection, screening x-ray mammography is recommended by health organizations worldwide and has been estimated to decrease breast cancer mortality by 20-40%. Nevertheless, significant false positive and false negative rates, as well as high interpretation costs, leave opportunities for improving quality and access. To address these limitations, there has been much recent interest in applying deep learning to mammography; however, obtaining large amounts of annotated data poses a challenge for training deep learning models for this purpose, as does ensuring generalization beyond the populations represented in the training dataset. Here, we present an annotation-efficient deep learning approach that 1) achieves state-of-the-art performance in mammogram classification, 2) successfully extends to digital breast tomosynthesis (DBT; "3D mammography"), 3) detects cancers in clinically-negative prior mammograms of cancer patients, 4) generalizes well to a population with low screening rates, and 5) outperforms five-out-of-five full-time breast imaging specialists by improving absolute sensitivity by an average of 14%. Our results demonstrate promise towards software that can improve the accuracy of and access to screening mammography worldwide.
Methods
The study combines multi-site mammography and DBT datasets with patient-level evaluation and a two-stage deep learning pipeline that progresses from patch classification to strongly and weakly supervised image-level training. A separate reader study evaluates five qualified breast-imaging radiologists on cancer, pre-index, and negative examinations.
- Evaluation design: Evaluation data were de-identified under IRB approval and divided at the patient level so exams from each patient remained in one cross-validation split.The reader study separately included 405 screening DM exams and five board-certified, MQSA-qualified breast-imaging radiologists reading cancer, pre-index, and negative cases across two sessions.
- Model training: Weakly supervised training performs binary cancer/no-cancer classification, while DBT training creates optimized slice images from the 2D model before weak supervision.The 2D stage used binary cross entropy, and DBT slices excluded the noisiest first and last 10% of each stack.
- Model training: Final predictions average three models across both image orientations, then average breast views and assign each study the higher of its two breast-level scores.This aggregation produces study-level scores for both 2D and 3D models.
Extended Data
Extended Data shows that the proposed model significantly outperformed recent models on both index and pre-index cancer exams and exceeded all tested reader combinations. Additional analyses evaluate performance under alternative case-composition definitions.
- Comparison to readers: The model outperformed every tested combination of two, three, four, or five readers on both index and pre-index cancer exams.Reader combinations averaged BIRADS scores, with sensitivity and specificity calculated using a threshold of 3.
- Comparison to recent work: The proposed model had significantly higher AUC than competing models on both index cancer exams and pre-index cancer exams.AUC differences used the DeLong method, with confidence intervals calculated from 10K bootstrap samples.
- Case composition: Additional analyses examined the proposed models under different case compositions, including a 12-month definition of index cancer exams.The main analysis used index exams acquired within 3 months before diagnosis and negative mammograms followed by another negative mammogram at the subsequent screen.