The accuracy of next-generation sequencing (NGS) for detecting tumor-specific mutations in plasma DNA is hindered by errors introduced during PCR/sequencing, base substitutions caused by DNA damage, and pre-existing mutations in normal cells that are present at a low frequency. Here, we performed NGS of genes related to pancreatic cancer (comprising 2.8 kb of genomic DNA) in plasma DNA (average 4.5 ng) using molecular barcodes. The average number of sequenced molecules was 900, and the sequencing depth per molecule was 100 or more. We also developed a bioinformatic variant filter, called CV78, to remove variants that were not considered to be tumor-specific, i.e., those that are either absent or occur at low frequencies in the Catalogue of Somatic Mutations in Cancer database. In a cohort comprising 57 pancreatic cancer patients and 12 healthy individuals, sequencing initially identified variants in 31 (54%) and 5 (42%), respectively, whereas after applying the CV78 filter, 19 (33%) and zero were variant-positive. In a validation cohort consisting of 86 patients with pancreatic cancer and 20 patients with intraductal papillary mucinous neoplasm (IPMN), 62 (72%) with pancreatic cancer patients and 10 (50%) IPMN patients were initially variant positive. After CV78 filtering, these values were reduced to 32 (37%) and 1 (5%), respectively. The variant allele frequency of filtered variants in plasma ranged from 0.25% to 76.1%. Therefore, combining NGS and molecular barcodes with subsequent filtering is likely to eliminate most non-tumor-specific mutations.