XIOSIS: Novel Deep Learning Method for Image-Guided Spacer Placement in Spacer-Enabled Pancreatic Cancer Radiotherapy
H Hooshangnejad 1,3*, J Lee 2,3, K Ding2,3, (1) Department of Biomedical Engineering, Johns Hopkins School of Medicine, Baltimore, MD, USA, (2) Department of Radiation Oncology and Molecular Radiation Sciences, Johns Hopkins School of Medicine, Baltimore, MD, USA, (3) Carnegie Center of Surgical Innovation, Johns Hopkins School of Medicine, Baltimore, MD, USA
Presentations
SU-J-202-5 (Sunday, 7/10/2022) 4:00 PM - 5:00 PM [Eastern Time (GMT-4)]
Room 202
Purpose: Spacer-enabled dose-escalated pancreatic cancer radiotherapy is shown to boost the effectiveness of treatment while sparing the radiosensitive duodenum. However, the complexity and uncertainty of endoscopic duodenal spacer placement have hindered its use. To address this, we developed a novel deep learning (DL) method for image-guided duodenal spacer injection and tracking. By providing intra-operative feedback, our method opens the way for robust spacer placement, enabling dose escalation, and improving the quality of life of more than 60,000 people diagnosed with devastating pancreatic cancer every year and tripling less than 10% 3-years survival rate.
Methods: We proposed and implemented the X-ray-guided intra-operative spacer injection system (XIOSIS), by developing a set of tools including (1) a realistic and patient-specific finite element-based model of duodenal spacer placement (FEMOSSA), (2) a Pix2pix X-ray out-painting network to extend the field of view (FOV) and overcome the shortcoming of small FOV portable C-arm X-ray images, and (3) an advanced fully convolutional generative adversarial network to synthesize the 3D intra-operative computed tomography (CT) augmented with duodenal spacer from three intra-operative 2D X-ray images.
Results: Our result suggests our system predicts the deformation of organs with, on average, less than 1 mm overlapped volume histogram, 2 mm nearest neighbor distance difference. The out-painting results showed an average pick-signal-to-noise ratio of 35.4±2.1 dB and a structural similarity index of 0.98±0.01. Finally, the synthesized CTs showed comparable spacer location and volume reconstruction with an averaged dice similarity coefficient of 0.82±0.05. Our simulation study showed that XIOSIS reduces the duodenal volume receiving 33Gy by 70 percent.
Conclusion: We developed XIOSIS, a novel DL method for image-guided spacer placement in spacer-enabled pancreatic cancer radiotherapy. XIOSIS offers a paradigm shift in the effectiveness of the complex endoscopic hydrogel spacer injection and dose-escalated radiotherapy for much-needed devastating pancreatic cancer and image-guided procedures in general.
Funding Support, Disclosures, and Conflict of Interest: Research Supported by National Cancer Institute R37CA229417
Keywords
Image Guidance, Treatment Techniques, Radiation Therapy
Taxonomy
TH- External Beam- Photons: intraoperative
Contact Email