MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel
  1. Home
  2. Meeting Program
  3. All
  4. Derive Patient-Specific Objectives Using a Neural Network Model Based On Overlapping Volume Histogram: A Step Towards Automated IMRT Planning in Nasopharyngeal Carcinoma Radiotherapy

Derive Patient-Specific Objectives Using a Neural Network Model Based On Overlapping Volume Histogram: A Step Towards Automated IMRT Planning in Nasopharyngeal Carcinoma Radiotherapy

Bai Penggang, Weng Xing, Dai Yitao, Chen Jihong, Chen Chuanben, Xu Yuanji, Qian Jiewei (1) Fujian Cancer Hospital, Fuzhou, China (2) Univeristy of South China, Hengyang, China(3) Fujian Cancer Hospital, Fuzhou, China (4) Fujian Cancer Hospital, Fuzhou, China (5) Fujian Cancer Hospital, Fuzhou, China (6) Fujian Cancer Hospital, Fuzhou, China (7) University of South China, Hengyang, China

Presentations

(Sunday, 7/14/2019)  

Room: ePoster Forums

Purpose: To validate a neural-network (NN) based automated IMRT planning technique for nasopharyngeal carcinoma (NPC) radiotherapy.

Methods: A total of 115 NPC patients treated with definitive step-and-shoot IMRT were involved in this study. Each patient has 25 items of overlap volume histogram (OVH) and a set of 21 planning objectives. Single layer NN was modeled in Python 3.6 with 275, 184 and 21 for the number of input, hidden and output nodes, respectively. Twenty-five NPC patients with their OVH were used as input for the NN model. Twenty-five patient-specific sets of objectives were derived using this model before their IMRT plans were automatically generated with one-shot optimization using Pinnacle scripts. The corresponding 25 manual IMRT plans were generated by one experienced medical physicist in our institute according to clinical requirements. Planning duration and dosimetric parameters were compared between automatic and manual plans.

Results: Compared to manual plans, automatic plans based on our method significantly reduced MU (685.24±58.89 vs. 721.36±63.36, P=0.004) and the planning duration (9.73±1.80 min vs. 57.10±6.35 min, P<0.001). Dosimetric parameters for PTV met the clinical requirements in both two types of plans. Though the automatic plans had a little worse homogeneity index for gtvtp and gtvnrp than the manual plans (1.095±0.018 vs.1.089±0.022, P=0.019; 1.042±0.011 vs. 1.036±0.009, P<0.001), the automatic plans had higher conformity index for ctv1p and ctv2p (0.757±0.049 vs. 0.728±0.048, 0.857±0.023 vs. 0.849±0.021; P<0.001). Compared with the manual plans, max dose for the right lens, left and right optic nerves in the automatic plans were significantly reduced by 0.32 Gy, 8.15 Gy, and 3.9 Gy, respectively (P<0.05).

Conclusion: Automatic method based on OVH using NN model can derive a robust set of planning objectives for a NPC patient. The automatic plans with one-shot optimization can also greatly reduce the planning time without compromising the plan quality.

Funding Support, Disclosures, and Conflict of Interest: Sponsored by Fujian Provincial Health Technology Project (Grant no. 2018-ZQN-19 and no. 2017-ZQN-15), the Fujian Provincial Natural Science Foundation (Grant no. 2015Y0010)

Keywords

Not Applicable / None Entered.

Taxonomy

IM/TH- Formal quality management tools: General (most aspects)

Contact Email

Additional Links