Room: Track 1
The use of deep learning techniques for detection and characterization of breast cancer in digital mammograms and tomosynthesis images is showing a lot of promise. There is an exponential growth of applications and publications in this field. However, at the same time, some specifics of the medical imaging field, such as the small size of available well-annotated data sets, substantial anatomical variabilities, and varying image quality, pose significant challenges. Different deep learning approaches for detection and characterization of breast cancer will be discussed. In addition, transfer learning techniques for deep learning networks, as a method to allow model generation from smaller data sets, will be presented. A deep learning based feature extraction and machine learning feature classification will also be presented. Some hybrid applications of deep learning approaches with other techniques will be discussed. The presentation will be concluded with discussion of potential pros and cons of different deep learning techniques when applied to detection and characterization of breast cancer on digital mammograms and tomosynthesis.
Implementation of these deep learning techniques in the clinic depend, first of all, on determining that they have achieved a certain level of performance. Therefore, it is important to evaluate and compare the performance of these methods against that of breast radiologists of varied experience levels, in different working environments (institutional-based screening, national screening programs, etc.), and with images acquired with systems from different vendors. Furthermore, it is crucial to determine what is the best way to implement these methods in the clinic. The notion that these systems will just replace breast radiologists is too simplistic, and not realistic, at least as a one-size-fits-all option. Various different implementation options are envisioned, all having a different impact on performance and human workload. Results of the studies on performance evaluation will be presented and discussed, and the questions they arise posed. The different implementation options that exist, or will become feasible in the future, will be reviewed, with their pros and cons discussed.
1. Understand different deep learning approaches for detection and characterization of breast cancer.
2. Learn about transfer learning, deep learning based feature extraction, and hybrid applications of deep learning approaches with other techniques.
3. Understand potential pros and cons of the different deep learning techniques when applied to detection and characterization of breast cancer on digital mammograms and tomosynthesis.
4. Understand the performance level that has been achieved with current deep learning-based methods for image interpretation of digital mammograms and tomosynthesis images.
5. Learn about different possible methods to introduce deep learning solutions for everyday clinical use in breast imaging.
Funding Support, Disclosures, and Conflict of Interest: ScreenPoint Medical is a spin-off company from the Department of the presenter. The presenter has no financial interest nor has received funding from ScreenPoint Medical.