Background:　Larval　zebrafish　phenotypes　serve　as　critical　research　indicators　in　fields　such　as　ecotoxicology　and　safety　assessment　since　phenotypic　defects　are　closely　related　to　alterations　of　underlying　pathway.　However,　identifying　these　defects　is　time-consuming　and　requires　specialized　knowledge.　Method:　We　proposed　a　deep　network　model　called　RECNet,　which　combines　attention　mechanisms　and　residual　structures.　In　terms　of　data　processing,　we　applied　the　mixup　data　augmentation　technique　and　accumulated　a　collection　of　6805　larval　zebrafish　phenotype　images,　mostly　generated　from　our　laboratory.　Our　proposed　model　was　deployed　to　execute　two　distinct　tasks,　including　a　four-classification　of　zebrafish　phenotypes　and　a　seven-classification　involving　mixed　labels　for　abnormalities.　Results:　In　the　four-class　classification　task,　the　RECNet　model　achieved　an　accuracy　of　0.949,　with　a　mean　area　under　the　curve　of　0.986　and　an　F1-score　of　0.966.　Through　interpretable　research,　attention　mechanisms　enable　the　model　to　focus　more　accurately　on　regions　of　interest.　In　the　mixed-label　seven-classification　task　for　anomalies,　our　model　achieved　an　accuracy　of　0.913　and　a　mean　average　precision　value　of　0.847　by　employing　the　weighted　loss　function　(DFBLoss).　Furthermore,　in　a　new　test　dataset,　the　RECNet　model　achieved　accuracy　rates　of　0.924　and　0.876　for　the　two　tasks,　respectively.　Our　RECNet　model　was　trained　by　orders　of　magnitude　larger　dataset　than　previous　studies　and　also　showed　better　accuracy　rates.　Conclusions:　Our　method　holds　promise　for　diverse　applications　within　zebrafish　laboratories　and　fields　such　as　toxicology,　providing　indispensable　support　to　scientific　research.　©　2025　Elsevier　Ltd