Predicting fluid intelligence in adolescence from structural MRI with deep learning methods

•To predict actual fluid intelligence from unsegmented MR images taking into account confounding image data and targeted scores.•The T1w MRI from the ABCD Neurocognitive Prediction Challenge (ABCD-NP-Challenge 2019) was utilised for prediction.•We show deep learning CNN models establish a weak but stable correlation between brain structure and raw fluid intelligence.•Spatially invariant image features contribute strongly in predicting uncorrected fluid IQ scores. Region specific information is necessary for predicting intelligence scores corrected for several covariates (age, sex, education etc.).