The document discusses the clinical applications of artificial intelligence (AI) in radiation oncology, highlighting various methods and predictive models using large datasets from electronic health records and medical imaging. It details results from different predictive approaches, including deep learning, showing accuracy in predicting therapeutic responses and survival outcomes. The presentation emphasizes the potential and challenges of integrating AI into clinical practices for personalized medicine.

1. Clinical applications of AI for
Radiation Oncology
Jean-Emmanuel Bibault, MD, PhD
Service d’oncologie radiothérapie - HEGP
INSERM UMRS 872, Equipe 22: Information Sciences to support
Personalized Medicine

2. Note
Please note that the slides of this presentation can
be shared on social media.

3. Conflict of interest
None

4. A few milestones
Source: Andrew Beam

5. The explosion of medical data

7. Many different methods within « AI »

9. Source: https://www.amax.com/blog/?p=804

10. A FEW EXAMPLES

11. TEXT DATA
Electronic Health Records

12. • 76214 patients
• Prediction of 78 diseases
Miotto et al, Scientific Reports, 2016

13. Miotto et al, Scientific Reports, 2016

14. 2 million adult and pediatric patients cared for at either the Stanford Hospital or the
Lucile Packard Children’s hospital between 1995 and 2014
Prediction of death (3-12 months)

16. MEDICAL IMAGING
DICOM

17. CheXnet
• Deep Neural Network (121
layers)
• Trained on 100 000 Chest X-
Rays
• 14 anomalies
• Saliency map
Rajpurkar P et al, arXiv, Dec 2017

18. Results
Rajpurkar P et al, arXiv, Dec 2017

19. Therapeutic response prediction

20. Results
Accuracy (IC 95%) Sensibility (IC 95%) Specificity (IC 95%) AUC (IC 95%)
Logistic Regression 69,5% (59,2% - 78,51%) 34,78% (16,38% - 57.27%) 80,56% (69.53% -
88.94%)
0,59 (0,458 – 0,686)
Support Vector
Machine
71,58% (61,4%-80,36%) 45,45% (24,39%-67,79%) 79,45% (68,38% -
88,02%)
0,62 (0.51 – 0.74)
Deep Learning 80% (70,54% - 87,51%) 68,2% (45,13% - 86,14%) 83,56% (73,05% - 91,21%) 0,72 (0.65 - 0.87)
Lambregts DMJ et al. 2011. « Diffusion-Weighted MRI for Selection of Complete Responders After Chemoradiation for Locally
Advanced Rectal Cancer: A Multicenter Study » Annals of Surgical Oncology
MRI 0-40% (0-53%) 92-98% (88-99%) 0,58-0,76 (0,47-0,86)
MRI + DWI 52-64% (85-99%) 89-97% (85-99%) 0,78-0,8 (0,69-0,91)
HUMAN PERFORMANCE?

21. Predict survival from
CT scans

23. PATHOLOGY

24. 270 patients
External validation on 129 other patients

25. AUC : 0.556 - 0.994

26. RADIATION ONCOLOGY

27. Automatic segmentation
S. Nikolov et al, ArXiv, 2018

29. • CNN
• 450 CT scans

31. Automatic treatment planning
Mahmood et al, ArXiv, 2018

32. BUT WE NEED TO REMAIN VERY CAREFUL…

35. jean-emmanuel.bibault@aphp.fr
twitter.com/jebibault
import tensorflow as tf
THANK YOU!