Proper Orthogonal Decomposition (POD) is a useful statistical tool for analyzing the cycle-to-cycle variation of internal combustion engine in-cylinder flow field. Given a set of flow fields (also known as snapshots) recorded over multiple engine cycles, the POD analysis optimally decomposes the snapshots into a series of flow patterns (known as POD modes) and corresponding coefficients of successively maximum flow kinetic energy content. These POD results are therefore strongly dependent on the kinetic energy content of the individual snapshots, which may vary over a wide range. However, there is as yet no algorithm in the literature to define, detect, and then remove outlier snapshots from the dataset in a systematic manner to ensure reliable POD results. In this paper, one such outlier detection algorithm is proposed: A snapshot is considered an outlier if it has excessively high kinetic energy content as well as relevance index (a measure of flow pattern similarity) versus a particular POD mode. An analytical expression is derived to relate these two parameters and provide the statistical and physical underpinnings for the outlier detection algorithm. The algorithm is applied to two sets of crank angle-resolved engine in-cylinder flow fields in the middle tumble plane that were obtained experimentally using high-speed particle image velocimetry techniques, and improvements in POD results obtained after the removal of outlier snapshots are discussed.