Diabetic　retinopathy　(DR)　is　an　ocular　manifestation　of　diabetes　and　the　leading　cause　of　visual　impairment　and　blindness　across　the　globe.　Early　detection　and　treatment　of　DR　can　salvage　from　visual　impairment.　The　manual　screening　of　DR　is　a　very　laborious　and　time-intensive　effort　and　heavily　dependent　on　professional　ophthalmologists.　In　addition,　the　subtle　distinction　among　various　retinal　biomarkers　and　different　grades　of　DR　makes　this　recognition　very　challenging.　To　address　the　aforementioned　problem,　deep　neural　networks　have　brought　many　revolutions　in　the　last　few　years.　In　this　study,　we　proposed　a　novel　two-stage　framework　for　automatic　DR　classification.　In　the　first　stage,　we　employed　two　distinct　U-Net　models　for　optic　disc　(OD)　and　blood　vessel　(BV)　segmentation　during　the　preprocessing.　In　the　second　stage,　the　enhanced　retinal　images　after　OD　and　BV　extraction　are　used　as　an　input　of　transfer　learning-based　model　VGGNet,　which　performs　DR　detection　by　identifying　retinal　biomarkers　such　as　microaneurysms　(MA),　haemorrhages　(HM),　and　exudates　(EX).　The　proposed　model　achieved　state-of-the-art　performance　with　an　average　accuracy　of　96.60%,　93.95%,　92.25%　evaluated　on　EyePACS-1,　Messidor-2,　and　DIARETDB0,　respectively.　Extensive　experiments　and　comparison　with　baseline　methods　demonstrate　that　the　effectiveness　of　the　proposed　approach.　©　2022　Informa　UK　Limited,　trading　as　Taylor　&　Francis　Group.