Deep learning dose prediction for IMRT of esophageal cancer: The effect of data quality and quantity on model performance

PurposeTo investigate the effect of data quality and quantity on performance deep learning (DL) models, for dose prediction intensity-modulated radiotherapy (IMRT) esophageal cancer.Material methodsTwo databases were used: a variable database (VarDB) with 56 clinical cases extracted retrospectively, including user-dependent variability in delineation planning, different machines beam configurations; homogenized (HomDB), created to reduce this by re-contouring re-planning all patients fixed class-solution protocol.Experiment 1 analysed variability, using 26 planned same machine setup (E26-VarDB versus E26-HomDB). Experiment 2 increased training set groups 10 (E16, E26, E36, E46, E56) both databases.Model evaluation metrics mean absolute error (MAE) selected dose-volume global MAE body voxels.ResultsFor 1, E26-HomDB reduced considered compared E26-VarDB (e.g. reduction 0.2 Gy D95-PTV, 1.2 Dmean-heart or 3.3% V5-lungs). For 2, increasing size slightly improved HomDB models decrease 0.13 E56-HomDB E26-HomDB), but VarDB increase 0.20 E56-VarDB E26-VarDB).ConclusionA small may suffice obtain good DL performance, provided that homogenous is used. Data reduces which further pronounced when set.