Diabetic　retinopathy　(DR)　is　a　common　diabetes　complication　that　can　cause　irreversible　blindness.　Deep　learning　models　have　been　developed　to　automatically　classify　the　severity　of　retinopathy.　However,　these　methods　face　challenges　like　a　lack　of　long-range　connections,　weak　interactions　between　images,　and　mismatches　between　lesion　details　and　receptive　fields,　leading　to　accuracy　issues.　In　our　research,　we　propose　a　deep　learning　model　with　three　main　aspects.　Firstly,　a　transformer　structure　is　incorporated　into　a　convolutional　neural　network　to　effectively　utilise　both　local　and　long-range　information.　Secondly,　the　disease　details　are　aggregated　from　multiple　images　before　applying　self-attention　to　improve　inter-image　interactions　and　reduce　overfitting.　Lastly,　an　attention-based　approach　is　proposed　to　filter　information　from　different　stages　of　feature　maps　and　adaptively　capture　lesion-related　details.　Our　experiments　achieved　a　5-class　accuracy　of　85.96%　on　the　APTOS　dataset　and　a　2-class　accuracy　of　95.33%　on　the　Messidor　dataset,　surpassing　recent　methods.