Non-ignorable missing data arise often in clinical trials. The VEST trial, a randomized, within-patient-controlled study, assessed the effect of an external scaffold for saphenous vein grafts on intimal hyperplasia (IH) one year after coronary artery bypass graft surgery. It was anticipated that approximately 13 % of grafts would be occluded and unsuitable for intravascular ultrasound, resulting in missing IH values at 1-year. Given graft occlusion is a negative outcome and higher IH is associated with occlusion, this missing data is non-ignorable. To address this, we developed a novel two-part method for the MNAR (missing not at random) scenario in the VEST trial. This method combines penalized multiple imputation with a modified Wilcoxon signed-rank test. We evaluated the approach's performance against alternatives in a series of simulation studies. The new method did not show type I error inflation. Under trial assumptions, it provided adequate power. However, if missing data exceeds 20 %, power decreases notably with the double penalization method due to underestimation of the treatment effect. When missing data is balanced between arms, penalized multiple imputation alone is more powerful and unbiased. Conversely, for unbalanced MNAR data, as might occur with a treatment effect on IH, the penalized multiple imputation with a modified Wilcoxon signed-rank test approach is more powerful. The VEST trial showed more occlusions than expected, balanced across arms, resulting in potential underestimation of the true treatment effect. However, given the potential for unbalanced missingness, this approach was suitable and could be applied in other settings with similar challenges.