Automatic　segmentation　of　left　atrial　cavity　and　scar　in　late　gadolinium　enhanced　magnetic　resonance　imaging　has　important　clinical　significance　for　the　diagnosis　of　atrial　fibrillation.　Owing　to　the　inferior　image　quality,　thin　walls,　surrounding　enhancement　regions,　and　complex　morphology　of　left　atrial　scars,　the　automatic　quantitative　analysis　of　them　is　extremely　challenging.　Either　manual　segmentation　of　the　left　atrial　cavity　or　the　atrial　scar　is　very　time-consuming　and　subjective　errors　may　occur.　In　this　work,　a　deep　neural　network　named　ResCEAUNet　has　been　developed　and　validated　for　automatic　segmentation　of　left　atrial　scars.　We　adopt　nnUNet　as　the　baseline.　To　enhance　segmentation　accuracy,　we　introduce　two　key　improvements　to　our　model:　the　lightweight　Convolutional　Block　Attention　Module　(CBAM)　and　the　edge　attention　module.　The　edge　attention　module　significantly　improves　the　model’s　ability　to　delineate　intricate　boundaries　of　the　atrial　wall　and　scar　tissue,　particularly　beneficial　for　thin　structures　like　the　left　atrium.　Simultaneously,　CBAM　sharpens　the　model’s　focus　on　relevant　features,　enabling　more　precise　localization　and　identification　of　scar　tissue　without　substantially　increasing　computational　complexity.　These　synergistic　enhancements　result　in　a　robust　and　efficient　segmentation　model,　demonstrating　its　effectiveness　by　achieving　a　Dice　score　of　0.6181　on　the　LAScarqs++　2024　validation　dataset.　©　The　Author(s),　under　exclusive　license　to　Springer　Nature　Switzerland　AG　2025.