Towards Clinically Effective AI for Screening Mammography

点击翻译按钮获取中文摘要

Abstract Breast cancer is the most common cancer among women and a leading cause of cancer mortality. Early detection of breast cancer can reduce mortality and morbidity, which has led to widespread mammography screening, recommended for women ages 50-74 on a yearly or bi-yearly basis. Reading the mammogram images to decide if cancer may be present is difficult due to the rarity of occurrence – in a screening population, 99.5% of women do not have cancer – and the visual challenge of finding what can be a very subtle abnormality on a complex background. This difficulty, combined with the high volume of mammograms – 39 million per year in the US – has led to a variety of proffered solutions including software known as computer-aided diagnosis (CAD). Despite early promise, such solutions have not fulfilled their potential in improving outcomes and are largely thought to increase interpretation times. Productivity is increasingly a concern due to the rapidly growing use of digital breast tomosynthesis (DBT or “3D” mammography), which has demonstrated higher cancer detection rates than traditional 2D mammography, but also takes much longer to interpret. As a potential solution, there has been significant interest in applying deep learning to mammography. Deep learning (DL) is a powerful field of machine learning which learns image features in an end-to-end fashion from data, and has been used to achieve human level performance on a number of imaging pattern recognition tasks. This proposal seeks to create DL-based software for mammography that can be effective in a clinical setting through (1) accurate and robust predictions on a diverse population of patients, (2) interpretable results from the DL model (no “black box” answers), and (3) applicability to both 2D mammography and DBT. In Phase I, the aim is to improve model performance by training on additional data and incorporating additional algorithmic advances. Phase I will conclude with a clinical reader study comparing performance of the software to radiologists. In Phase II, the aim is to improve the clinical effectiveness of the software by automating quality detection, incorporating prior exams into the model, and expanding the training dataset to ensure results generalize to any woman eligible for screening mammography. These improvements will apply to both 2D and DBT. Achieving the desired performance levels will enable a product that will improve productivity for radiologists and ensure consistent and accurate interpretations for patients. Success in this project would be a large step towards translating state-of- the-art artificial intelligence to clinically effective tools for screening mammography.