1. 1. Key concepts
   1. 1. What is Deep Learning (DL)?
      1. Aims to imitate human thought process via data analysis and algorithms called artificial neural networks
   2. 2. How is Dynamic Contrast-Enhanced Magnetic Resonance Imaging (DCE-MRI) used for breast cancer diagnostics?
      1. 1. Highly sensitive (80%) test used for detecting breast cancer
      2. 2. DCE-MRI results can be classified into 6 distinct categories within a BI-RADS system.
         1. Type 1 = Negative
         2. Type 2 = Benign finding
         3. Type 3 = Probably benign finding
         4. Type 4 = Suspicious abnormality
            1. MUST DO BIOPSY
         5. Type 5 = Malignant
         6. Type 6 = Already proven to have been malignant
2. 2. What is the current issue?
   1. Positive Predictive Value (PPV) of biopsy results of women diagnosed with BI-RADS type 4 range from 19.6% to 35.7%
      1. In other words, for every malignant biopsy, 2-4 other biopsy findings are benign
         1. These benign results
            1. Cause stress & patient workup!!
            2. Result in unnecessary biopsies!!
3. 3. How can DL solve this current issue?
   1. 1. Improve the MRI specificity
      1. Reduce the number of false positives
      2. Minimize unnecessary biopsies
      3. Reduce patient workup
   2. 2. Improve clinical impact by facilitating the translation from results to intervention
4. 4. Testing Methods
   1. 1. Evaluating DL-MRI performance by comparing it with radiologists' diagnostics using AUROC curves
   2. 2. Using Decision Analysis Curve (DCA) to determine clinical utility impact on decision-making
      1. Example: Is it possible to downgrade a patient from BI-RADS 4 to BI-RADS 3 and, thus, reduce unnecessary biopsies?
5. 4.1 Evaluating DL-MRI performance using AUROC curves
   1. TPR (sensitivity) vs FPR (1-specificity)
   2. DL-MRI is either just as good or better than radiologists' diagnostics.
6. 4.2 Using DCA to determine clinical utility impact
   1. Is it possible to downgrade a patient from BI-RADS 4 to BI-RADS 3?
      1. Standard net benefit
         1. % of unnecessary biopsies avoided
      2. Decision threshold
         1. Variable according to threshold probabilities and to doctor/patient preferences
      3. Even at low decision thresholds, DL-MRI proves to be much more effective than a biopsy all intervention
7. 5. Limitations to this study?
   1. 1. No investigation on how DL-MRI makes its decisions.
      1. DL-MRI doesn't specify what specific part of the image influenced their prediction
   2. 2. No cost-effective analysis on benefits of DL-MRI
8. 6. Conclusion
   1. 1. DL-MRI possesses diagnostic accuracy equivalent to radiologist experts
   2. 2. Improves clinical impact utility via personalized patient management