radiotherapy Dr Alex Zwanenburg 06122021 Why should we personalise radiotherapy amp chemoradiotherapy after 5 years amp chemoradiotherapy after 5 years John Doe Richard Roe ID: 930628
Download Presentation The PPT/PDF document "ML for personalised" is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
Slide1
ML for personalised radiotherapyDr. Alex Zwanenburg
06.12.2021
Slide2Why should we personalise radiotherapy?
&
chemoradiotherapy
✓
after 5 years
&
chemoradiotherapy
✗
after 5 years
John Doe
Richard Roe
Slide3Why should we personalise radiotherapy?3
Each patient responds differently to radiotherapy.
How can we provide the best treatment for each patient?
Optimal patient-specific dose for
tumour
control.
Prevent patient-specific side effects due to radiation.Ensure correct treatment delivery.treatment #1treatment #2
treatment #3
Slide4Data sources4
Imaging
Blood
Cytotoxicity
Demographics
Genomics
PathologyProteomicsPlanned doseAnd more …
Machine data
Lambin
et al. (2012) 10.1016/j.ejca.2011.11.036.
Slide51. Multicentre clinical studies5
Analysis of
clinical
,
imaging
and
transcriptome data within German Cancer Consortium (DKTK):Head and neck squamous cell carcinoma (HNSCC)Rectal cancerGlioblastomaTypical endpoints: tumour control, overall survivalAim: Identify patient risk groups for future treatment individualization
Slide61. Example imaging study6
Aim: Compare deep learning architectures for
time-to-event
endpoints
with
clinical imaging
Transfer learning2D (VGG-like)3D (Hosny et al.)Auto-encoderStarke et al. (2021) 10.1038/s41598-020-70542-9
Slide71. Example imaging study7
Starke et al. Scientific Reports 2021, 10: 15625.
3D CNN (CI=0.69) 2D CNN (0.62) Transfer learning (0.63)
CT-based deep learning predicts recurrences after treatment
Best results for 3D networks
Loco-regional control
Slide82. Proton or photon therapy?8
Patient A
Model-based approach
?
Aim
:
Support therapy decision using prognostic models for treatment related side effectsLangendijk et al (2013) 10.1016/j.radonc.2013.05.007; Dutz et al. (2020) 10.1016/j.radonc.2019.12.024; Dutz et al. (2021) 10.1016/j.radonc.2021.04.008NTCP = Normal tissue complication probabilityPatient B
Slide93. Relative biological effectiveness of protons9
D
physical
D
physical
RBEvariableLinear energy transferProtons are more effective than photons, Clinics: relative biological effectiveness (RBE) = 1,1 Reality: RBW ≠ 1,1Aim: Predict RBE and side effects, include in treatment planningMonte-Carlo simulation NTCP model
Dosis
LET
NTCPDLETLühr et al, Radiother Oncol 2018; Eulitz et al., Phys Med Biol 2019, Acta Oncol 2019, Radiother Oncol 2020
Slide103. Range verification of proton therapy
Geplante Dosis
Applied dose
Messungen:
Applizierte Dosis
Aim
: Automated detection of relevant range deviations
Richter et al.,
Radiother
Oncol 2016; Nenoff et al., Radiother Oncol 2017; Khamfongkhruea et al., Med Phys 202010
Slide114. MRI to pseudo-CT11
Aim
:
allow
for MRI-only radiotherapy workflowsKazemifar et al. (2020) 10.1002/acm2.12856
Slide125. Tools12
Deep learning
Machine learning
Slide135. Tools13
Suitable for classification, regression and survival endpoints.
Perform model development, evaluation and explanation steps
automatically
in an
end-to-end
fashion.Enclose all relevant information in models.Advantages:Makes modelling easier and less error-prone.Relevant reporting information is available.Minimal user interaction less bias.Zwanenburg et al. (2021) https://github.com/alexzwanenburg/familiar/ *Currently in pre-release; contact for access
Slide14Thank you for your attention!
Dr Alex Zwanenburg
alexander.zwanenburg@nct-dresden.de
alex.zwanenburg@oncoray.de