The advent of digital breast tomosynthesis (DBT) or ?3D mammography? brings promise of improved sensitivity and specificity. Even as DBT undergoes rapid clinical adoption, this new imaging modality poses new challenges to radiologists due to many differences in image appearance compared to conventional mammography. Recent studies and clinical experience show the urgent need for improved training in DBT. These changes necessitate dedicated training for both radiology residents and board-certified radiologists. The currently accepted training paradigm (i.e., the ?one size fits all? approach) cannot effectively accommodate the individual needs of each radiologist-in-training. We propose to address this shortcoming. This project seeks to develop a new paradigm of adaptive, computer-enhanced training in the interpretation of DBT. Using predictive algorithms similar to the recommendation systems of companies like Amazon and Netflix, our system will create a training environment tailored to the needs of each individual radiologist. We have shown that the radiologists-in-training do not make errors in a random manner; instead their error making demonstrates specific patterns. Even more importantly, those errors can be predicted from each individual?s previous interpretations. We propose to use machine learning-based user modeling as well as collaborative filtering algorithms to identify the strengths and weaknesses of each individual. Then, we will adapt the training protocol to the individual needs of the trainee in order to accelerate the training and improve his/her diagnostic skills in the areas that need improvement.
The aims are: (1) develop user models that predict the difficulty of a DBT case based its image features and trainee performance, (2) implement an educational system that presents training cases that focus on the unique needs of each individual, and (3) conduct a reader study to validate the improved effectiveness of the proposed adaptive system. This has work has the following key innovations. We propose the first ever computer-aided education system for residents and radiologists that is adaptive and individualized. We leverage cutting edge concepts from the fields of user modeling and collaborative filtering, which our lab has introduced to the field of medical imaging. This work addresses training limitations for the newly adopted modality of DBT. This research has very high clinical significance: Improving training in the interpretation of breast images will ensure the greatest benefit of breast cancer screening. The proposed work, while developing a fully functional DBT education system, will contribute to the foundation for a more general, new paradigm of improved individualized education that extends to other image interpretation tasks and imaging modalities. This work is a collaboration between computer vision and machine learning experts, breast imaging radiologists, director of Radiology residency, a biostatistician, and a human factors expert. Our technology will be disseminated through a partnership with IMAIOS, a leading provider of online radiology education.
Digital breast tomosynthesis is becoming a standard practice in breast cancer screening and there is a pressing need for effective education in this new modality. In this study we propose to develop a computer system for education in digital breast tomosynthesis that adapts itself to each trainee and to find an individualized training protocol that fits best the trainee?s needs. Such system could improve dramatically the time efficiency of training in digital breast tomosynthesis images and improve diagnostic performance of the trainees. This is likely to improve the overall effectiveness of breast cancer screening and, given high importance of breast cancer screening, decrease mortality rate of the disease.
|Harowicz, Michael R; Robinson, Timothy J; Dinan, Michaela A et al. (2017) Algorithms for prediction of the Oncotype DX recurrence score using clinicopathologic data: a review and comparison using an independent dataset. Breast Cancer Res Treat 162:1-10|