This　paper　presents　a　semantic　segmentation　method　that　can　distinguish　six　different　types　of　intracranial　hemorrhage　and　calculate　the　amount　of　blood　loss.　The　major　challenge　of　medical　image　segmentation　are　the　lack　of　enough　data　due　to　the　difficulty　of　data　collection　and　labeling.　In　this　paper,　we　propose　to　adopt　a　pretrained　U-Net　model　with　fine　tuning　to　solve　this　problem.　The　best　final　test　accuracy　can　reach　94.1%　which　is　10.5%　higher　than　the　model　training　from　scratch,　proving　its　advantages　in　dealing　with　relatively　complex　datasets　with　a　small　amount　of　data,　and　the　success　of　the　proposed　segmentation　method.