ESTRO 2020 Abstract Book

S1093 ESTRO 2020

Monday 30 November

Session: Optimisation, verification & prediction

OC-1040 Computer aided brachytherapy: assisting the practice of prostate brachytherapy with machine learning J. Sanders 1 , A. Venkatesan 2 , J. Davis 3 , R. Kudchadker 4 , C. Tang 5 , T. Bruno 5 , J. Ma 1 , S. Frank 5 1 University of Texas MD Anderson Cancer Center, Imaging Physics, Houston, USA ; 2 University of Texas MD Anderson Cancer Center, Diagnostic Radiology, Houston, USA ; 3 University of Texas MD Anderson Cancer Center, Urology, Houston, USA ; 4 University of Texas MD Anderson Cancer Center, Radiation Physics, Houston, USA ; 5 University of Texas MD Anderson Cancer Center, Radiation Oncology, Houston, USA Purpose or Objective To demonstrate how machine learning can be incorporated into an MRI-only low-dose-rate (LDR) prostate brachytherapy workflow to facilitate index lesion identification, treatment planning, and postimplant assessment. Material and Methods Patients undergoing prostate cancer treatment with LDR brachytherapy were evaluated with an MRI-only workflow, which included a multiparametric MRI (mpMRI) for diagnosis (T1w, T2w, DWI, and DCE-MRI), T2w anatomical imaging for treatment planning and simulation, and T2/T1w anatomical imaging for postimplant assessment (Figure 1). Machine learning algorithms that operated on the clinical MRIs at each step in the treatment chain were developed. Fully convolutional networks (FCNs) were constructed for dominant lesion identification and segmentation on pre- treatment diagnostic mpMRIs. A multi-tasking neural network based on FCNs was constructed for anatomy contouring in both pre- and post-implant MRI. Based on the preimplant segmentation, auto-treatment planning can be performed with a dictionary based algorithm implemented in a commercial treatment planning system (TPS) (MIM Software, Cleveland, OH). A two-stage sliding-window convolutional neural network (CNN) was previously developed for post-implant seed identification and localization. Predictions from the post-implant seed identification CNN and post-implant anatomy contouring FCN were combined for automated post-implant dosimetry. Some of the algorithms were implemented into a commercial TPS using a vendor provided software development kit (Figure 2). Results Predictions on the diagnostic mpMRIs, preimplant T2w, and postimplant T2/T1w MRIs took approximately 1 min, 1 min, and 3 mins per patient, respectively. Overall, dominant lesion identification at a confidence threshold of ≥50% was achieved with 72.4% sensitivity and 35.8% precision. Dice coefficients between the manual and automated masks of the prostate, seminal vesicles, rectum, and bladder in the preimplant T2w MRI were 86.9%, 62.3%, 93.7%, and 93.7%, respectively. For the postimplant T2/T1w MRI, the Dice coefficients of these organs were 89.4%, 79.0%, 91.0%, and 96.3%, respectively. Precision and recall of the radioactive seeds at a confidence threshold of ≥50% on the postimplant T2/T1w MRI were 96.1% and 96.4%, respectively.

Conclusion The availability of open-source machine learning software frameworks, general purpose graphics processing units, and modern computer vision techniques have enabled the automation of many routine tasks in clinical workflows, requiring a fraction of the time by a human to make decisions. We have investigated development and utilization of these techniques in the context of an MRI- only LDR prostate brachytherapy workflow. Incorporating these algorithms into a commercial TPS has the potential to substantially reduce the amount of time required to manage MRI-based LDR prostate cancer treatments, and potentially expand the access to precision LDR prostate brachytherapy. OC-1041 In-vitro determination of radiobiological parameter values used in cervical cancer brachytherapy B. Chow 1 , K. Nanda 2 , B. Warkentin 1 , F. Huang 1 , A. Gamper 1 , G. Menon 1 1 University of Alberta, Oncology, Edmonton, Canada ; 2 University of Alberta, Cell Biology, Edmonton, Canada Purpose or Objective Brachytherapy (BT) boosts used to treat cervical cancer, given after external beam radiotherapy (EBRT), can be delivered with either high-dose-rate (HDR) or pulsed-dose- rate (PDR). The boost prescription utilizes radiobiological dose (units of EQD2, EQuivalent Dose in 2 Gy fractions) and uses parameters α/β , a characterization of the sensitivity of the tissue to fractionation, and T 1/2 , the repair rate of sublethal DNA damage. Conventionally, α/β = 10 Gy for tumours and T 1/2 = 1.5 hours are used for cervical cancer BT dose calculation. With current recommendation to escalate the total prescription to at least 90 Gy EQD2, a better understanding of radiobiological parameter values is required to reduce dose uncertainties. This study was conducted to determine α/β and T 1/2 values through in vitro experiments using cervical cancer cell lines and clinically relevant BT treatment schedules. Material and Methods Cervical cancer cell lines (C-33A, CaSki, SiHa) were irradiated with different dose rates and fractionation schedules, using a Cs-137 irradiator and clinical Ir-192 HDR