Knowledge-based automated radiation therapy treatment planning utilizing dose prediction with a 2.5D-U-Net

Inverse treatment planning (ITP) in intensity modulated radiation therapy (IMRT) involves the specification of dose constraints as objective functions. These are manually and iteratively adjusted during planning to achieve patient-specific optimal outcomes. We propose a knowledge-based treatment planning (KBP) solution using deep learning to automate ITP.

KBP is tested for prostate and breast volumetric modulated arc therapy. A 2.5D-U-Net is optimized for dose prediction using 72/66 treatment plans obtained from manual planning (MP). The dose prediction model is then used to derive personalized optimization parameters to control ITP for 12 test patient datasets each (60/50Gy prescription dose). For inference, MP and KBP outcomes are compared using DVH metrics, plan quality metric (PQM) and blinded expert rating.

Median differences (MD) between MP and KBP are maximum 0.5Gy in absolute values for prostate organs at risk (OAR) mean doses. MD in V_56Gy is 0.4cc for rectum and bladder. For breast treatment, absolute MD values for OAR mean doses are maximum 0.4Gy. MD in V_20Gy of the ipsilateral lung is -1.0%. MD in PQM is 0.2% and 2.9% for prostate and breast, respectively. According to expert rating, in none of the test cases KBP provides a markedly worse plan than MP.

We developed a KBP solution for ITP of IMRT that is able to automatically generate treatment plans with similar quality when compared to MP. This has the potential to facilitate and accelerate ITP while ensuring high treatment plan quality.