Objective:　Diabetic　retinopathy　(DR)　is　a　common　complication　caused　by　diabetic　and　can　lead　to　severe　visual　impairment　or　even　blindness.　With　the　rise　of　DR　population,　clinicians　demand　to　adopt　an　automated　diagnosis　for　reducing　some　burden.　Automated　DR　classification　may　improve　efficiency　and　accuracy,　and　reduce　the　burden　of　clinicians.　Methods:　A　full　convolution　spatial　attention　module　(FCSAM)　and　CS-ResNet-101　model　are　proposed　for　DR　classification.　In　addition,　an　image　enhancement　algorithm　is　proposed　to　improve　the　quality　of　DR　images.　Furthermore,　the　cross-entropy　loss　function　is　improved　to　reduce　the　overfitting.　The　EyePACS　dataset　is　utilized　as　training　and　testing,　and　the　APTOS　2019　dataset　is　used　as　external　validation.　Results:　By　comparing　evaluation　metrics　of　DR　classification　among　different　models,　the　findings　are　as　follows:　(1)　Accuracy,　specificity,　sensitivity　and　F1　score　of　the　CS-ResNet-101　model　with　the　image　enhancement　algorithm　and　the　improved　loss　function　reach　98.1%,　99.6%,　98.1%　and　98.1%,　respectively,　and　it　converges　faster　than　other　models.　(2)　The　CS-ResNet-101　model　exhibits　excellent　generalization　performance　on　the　external　validation.　Conclusion:　The　above　experimental　results　indicate　that　the　proposed　model　with　the　image　enhancement　algorithm　and　the　improved　loss　function　is　effective　and　advanced　in　terms　of　DR　classification.　Advances　in　knowledge:　This　work　proposes　an　automated　DR　classification　method　based　on　deep　learning　and　an　image　enhancement　algorithm　for　DR　image.