Click here to


Are you sure ?

Yes, do it No, cancel

Prediction of Dosimetric Variation in Dominant Intraprostatic Lesion with Simultaneous Integrated Boost for Pencil Beam Proton Therapy

C Chang*, D Bohannon, K Stiles, A Stanforth, S Tian, Y Wang, L Lin, P Patel, T Liu, X Yang, J Zhou, Emory Proton Therapy Center, Atlanta, GA


(Sunday, 7/12/2020)   [Eastern Time (GMT-4)]

Room: AAPM ePoster Library

Purpose: beam proton therapy can deliver simultaneous integrated boost (SIB) to the dominant intraprostatic lesion (DIL) which may improve prostate radiotherapy efficacy. Proton SIB to DIL requires high precision in dose coverage. However, DIL dose coverage is sensitive to patients’ anatomical changes such as bladder and rectum filling. We propose a deep learning-based online prediction method to estimate DIL dose coverage deviation due to patients’ anatomical changes detected by pretreatment cone-beam CT (CBCT).

Methods: neural networks (FNNs) were used with data from twenty-five prostate patients with SIB plans (CTV V100%>99% and DIL V98Gy>95%) and 122 CBCT images. DILs were contoured based on co-registered multiparametric MRI images. Anatomical changes such as bladder and rectum volumes, femur shift, and the distance from DIL to bladder and rectum were calculated from first-week daily CBCT images and used as FNN inputs to predict if DIL V98Gy drop > 10% and bladder/rectum maximum dose is > 80Gy. FNN’s performance was evaluated with ROC curves.

Results: medians for DIL V98Gy, bladder and rectum maximum dose are 98.5%, 74.5 Gy, and 74.9 Gy from 122 cases. The overall accuracies of FNN in predicting DIL V98Gy drop > 10% (29 out of 93), bladder and rectum maximum dose > 80Gy (10 out of 112 and 16 out of 106) are 92.3%, 100%, and 92.3%, and the area under the ROC are 0.85, 0.99, and 0.97, respectively.

Conclusion: developed a FNNs method to predict dosimetric changes for DIL with SIB, and this method can inform physicians in terms of the impact of anatomical changes to DIL coverage. A substantial amount of data is essential to increase the robustness of FNNs, which can be adapted to broader scenarios.


Not Applicable / None Entered.


Not Applicable / None Entered.

Contact Email