The　human　ears　are　very　precise　and　contain　more　than　30　anatomical　structures.　U-HRCT　is　a　dedicated　Ultra-High　Resolution　CT　for　otology,　which　usually　generates　vast　amount　of　imaging　data.　Supervised　image　segmentation　can　achieve　high　segmentation　performance,　but　requires　a　large　number　of　pixel-level　manual　annotations.　For　U-HRCT　images,　manual　annotation　is　very　time-consuming　and　labor-intensive.　So　it　is　impractical　to　provide　large-scale　annotated　samples　for　each　ear　structure　segmentation　tasks.　In　this　paper,　a　few　shot　U-HRCT　ear　structure　segmentation　method　is　proposed　by　meta-knowledge　learning　from　few　existing　relevant　tasks.　Our　important　insight　is　to　learn　the　meta-prior　knowledge　from　the　existing　few　relevant　U-HRCT　image　ear　structure　segmentation　tasks,　which　is　then　transferred　to　a　new　ear　structure　segmentation　task　to　guide　its　learning　and　training.　Therefore,　the　network　can　quickly　adapt　to　the　new　task　and　achieve　a　satisfactory　segmentation　accuracy　with　only　few　labeled　data　of　the　new　task.　Using　CE-Net　as　the　basic　segmentation　network,　we　verified　the　proposed　method　on　the　self-established　U-HRCT　ear　structure　segmentation　dataset.　The　experimental　results　show　that,　through　learning　the　meat-knowledge　from　the　existing　malleus,　incus,　and　stapes　segmentation　tasks,　for　the　new　inner　ear　labyrinth　segmentation　task　with　only　4　labeled　cases,　its　DSC　index　can　reach　90.70%,　which　greatly　reduces　the　manual　annotation　workload　of　doctors.　©　2024　ACM.