Aiming　at　the　problems　of　difficult　segmentation　and　inaccurate　segmentation　of　small　lesions　in　lung　Computed　Tomography　(CT)　images,　a　multi-scale　feature　fusion　deep　learning　prediction　model　that　conforms　to　the　observation　of　things　by　the　human　eye　is　proposed,　and　specific　performance　indicators　and　visualization　results　are　given.　The　multi-scale　feature　fusion　mechanism　is　used　to　effectively　capture　the　long-distance　characteristics　of　lesions.　A　joint　loss　function　is　proposed　to　make　the　training　smoother　and　further　improve　the　segmentation　performance.　The　analysis　is　verified　in　the　segmentation　test　set,　and　the　results　show　that　the　overlap　between　the　proposed　model　segmentation　results　and　the　real　results　has　a　Dice　value　of　83.29%,　a　sensitivity　Sen　of　82.66%,　a　cross-union　ratio　IoU　of　73.15%,　and　a　specificity　Spec　of　99.82%,　which　can　better　segment　the　lesion　region　compared　with　the　existing　algorithms.　Therefore,　using　the　proposed　improved　model　to　predict　lung　CT　image　lesions　can　be　used　clinically　for　doctors　to　diagnose　diseases　faster.　©　2024　IEEE.