Click here to

Session: Deep Learning Response Prediction, Diagnosis, and Modeling [Return to Session]

Predict Voxel-Wise Active Bone Marrow Loss of Anal Cancer Patients Treated with Chemoradiation Using 3D U-Net Deep Learning of Post-Radiotherapy PET Images

Y Yue*, Y Le, O Ishaq, T Lautenschlaeger, Indiana University School of Medicine, Indianapolis, Indiana

Presentations

WE-C930-IePD-F6-6 (Wednesday, 7/13/2022) 9:30 AM - 10:00 AM [Eastern Time (GMT-4)]

Exhibit Hall | Forum 6

Purpose: Irradiation-induced hematologic toxicities (HT) of anal cancer patients during chemoradiotherapy can be evaluated by the loss of active bone marrow (ABM) measured from PET images. To predict voxel-wise ABM loss before treatment starts, we propose to predict the post-radiotherapy PET images using the pre-treatment PET images and planning information.

Methods: Fifty-five anal cancer patients treated with chemoradiotherapy were studied. All patients had 18FDG-PET/CT images 2-weeks prior to(pre-RT), 6-months following treatment(post-RT). Pelvic ABM was characterized in PET images as the volume having standard uptake value (SUV) greater than background. The bladder uptake of post-RT PET was reduced to the mean liver uptake to reduce the bias. The prediction of post-RT PET image was constructed by the 3D U-Net deep learning with seven-channel inputs: PTV, V40, pelvic bone, bladder, Pre-RT PET, planning CT, and planned dose. The performance of the prediction model was evaluated by the voxel-based joint histograms. The mean SUV, ABM and ABM loss of pelvic bone were calculated. The Analysis-of-variance (ANVOA) was used to identify statistical difference between the predicted post-RT PET and ground truth images.

Results: Forty patients were randomly selected as training dataset, and fifteen patients as test dataset. The predicted post-RT PET images were compared with ground truth by the joint histogram with voxel-wise correlation (mean correlation coefficient, 85.15%+/-9.48). The mean SUV of pelvic bone marrow was 0.89+/-0.18 for predicted images, and 0.92+/-0.23 for ground truths with p=0.372. Average ratio of pelvic ABM to total bone marrow was 34.8%+/-10.1% for predicted images, and 35.2%+/-13.2% for ground truths. The mean ABM loss was 21.1% for predicted images, and 20.7% for ground truths with p=0.787.

Conclusion: Our approach can effectively predict voxel-wise post-RT metabolic response using pre-RT and treatment information. Prediction of ABM loss provides a feasible approach to implement response-driven adaptive radiotherapy for anal cancer.

Keywords

PET, Dose Response, Image-guided Therapy

Taxonomy

TH- Response Assessment: Imaging-based: PET

Contact Email

Share: