Background: Precision medicine for multiple sclerosis (MS) involves choosing a treatment that best balances efficacy and disadvantages such as side effects, cost, and inconvenience, based on an individual’s unique characteristics. Machine learning can be used to model the relationship between a baseline brain MRI and future new and enlarging T2 (NE-T2) lesion count to provide personalized treatment recommendations. Methods: We present a multi-head, deep neural network for making individualized treatment decisions from baseline MRI and clinical information which (a) predicts future NE-T2 lesion counts on multiple treatments and (b) estimates the conditional average treatment effect (CATE), as defined by the predicted suppression of NE-T2 lesions, between different treatment options and placebo. We validate our model on a dataset pertaining to 1817 patients from four randomized clinical trials. Results: Our model predicts favorable outcomes (< 3 NE-T2 at follow-up) with average precision 0.780-0.994 across 5 different treatment arms. It correctly identifies subgroups with different treatment effect sizes and provides treatment recommendations that improve lesion suppression while limiting the need for high efficacy treatments. Conclusions: Our framework provides accurate predictions for future NE-T2 lesion counts and personalized treatment recommendations that improve outcomes while accounting for the disadvantages of different treatment options.