Gastric　cancer　is　one　of　the　highest　mortality　cancers　in　the　world.　At　present,　pathological　diagnosis　is　the　gold　standard　for　gastric　cancer　diagnosis.　Deep　Learning　technique　has　been　widely　used　in　pathological　diagnosis　using　medical　images.　The　differentiation　of　gastric　cancer　cells　is　one　of　the　important　contents　in　the　pathology　report.　The　degree　of　tumor　differentiation　reflects　the　degree　of　cell　maturity　and　the　degree　of　malignancy　of　the　tumor.　It　is　of　great　significance　to　carry　out　early　treatment　and　prognostic　diagnosis.　However,　there　are　few　related　researches　on　the　calculation　of　gastric　cancer　cell　differentiation　degree　in　medical　imaging,　since　it　is　difficult　to　accurately　segment　the　ROI　(region　of　interest)　of　gastric　cancer　cells.　In　this　paper,　we　propose　an　improved　Mask　R-CNN　framework　that　called　Mask　Attention　R-CNN　to　predict　the　differentiation　degree　of　gastric　cancer　cells.　The　framework　contains　a　mask　attention　block　(MAB)　that　can　calculate　the　Intersection　over　Union　(IoU)　of　the　predicted　gastric　cancer　cell　mask　and　mask　feature　map　to　enhance　the　mask　quality.　We　also　add　a　scale　jittering　strategy　which　can　learn　detailed　information　from　images.　As　far　as　we　know,　this　is　the　first　study　to　apply　the　instance　segmentation　framework　to　the　task　of　predicting　differentiation　degree　of　gastric　cancer　cells.　The　mask　branch　effectively　improves　the　accuracy　of　the　instance　mask　of　gastric　cancer　cells.　Both　the　location　and　classification　accuracies　of　gastric　cancer　cells　are　improved　by　our　model.　This　can　provide　pathologists　with　auxiliary　diagnosis　opinions,　which　has　certain　clinical　application　value.　©　2021　IEEE.