Multicentric Study for External Validation of a Deep Learning Model for Mammographic Breast Density Categorization

Multicentric Study for External Validation of a Deep Learning Model for Mammographic Breast Density Categorization

The correct categorization of breast density is essential to adapt the diagnostic examination to the needs of each patient. Assessment of breast density is performed visually by radiologists. Some authors have detected that this method involves considerable intra and interobserver variability. On the other hand, automated systems for measuring breast density are becoming more and more frequent. Machine learning is a domain of Artificial Intelligence, which comprises the process of developing systems with the ability to learn and make predictions using data. These systems are designed to aid healthcare professional decision making. In the present work, the multicenter study of external validation of a tool based on deep learning for the categorization of mammographic breast density is proposed.

The correct categorization of breast density is essential to adapt the diagnostic examination to the needs of each patient. Assessment of breast density is performed visually by radiologists. Some authors have detected that this method involves considerable intra and interobserver variability. On the other hand, automated systems for measuring breast density are becoming more and more frequent. Consequently, in clinical practice, breast density is reported from the assessment carried out by specialists with the support of these systems. But there are few studies about the use, concordance and perception of usefulness of professionals on these tools. A study carried out at the Hospital Italiano de Buenos Aires reported a moderate to almost perfect inter- and intra-observer agreement among radiologists and a moderate concordance between the categorization carried out by experts and that carried out by commercial software of a digital mammography machine. Machine learning is a domain of Artificial Intelligence, which comprises the process of developing systems with the ability to learn and make predictions using data. Once a system designed to aid healthcare professional decision making is developed, it must be validated. In 2019, an internal validation of a tool based on deep learning techniques was carried out for the automatic categorization of mammographic breast density. The tool reached a very good interobserver agreement, kappa = 0.64 (95% CI 0.58-0.69), when compared with the performance of the professionals. It reached a sensitivity of 83.2 (CI: 76.9-88.3) and a specificity of 88.4 (83.9-92.0.) In the present work, the multicenter study of external validation of a tool based on deep learning for the categorization of mammographic breast density is proposed. The evaluation of this tool will be carried out in two external institutions: Hospital Alemán and Fundación Científica del Sur.

Inclusion Criteria:

• Mammograms included in the study should meet the following criteria:

  • Female patients of 40 years of age or more.
  • To have at least one screening mammography exam performed at Saint John's
  • Cancer Institute during the study period. These exams will be included regardless of the brand of the mammography equipment.
  • Mammograms should be performed with digital equipment.

Exclusion Criteria:

• Mammograms with the following criteria will be excluded from the study:

  • Patients with gigantomastia, defined by the need for more than one image of each mammographic view (mediolateral oblique and craniocaudal) to evaluate the entire breast volume.
  • Patients with breast implants.
  • Patients with a history of breast surgery.