Automated segmentation of dorsal muscle group in abdominal

基於人 智慧之自動化標示骨骼肌群 Automated segmentation of dorsal muscle group in abdominal CT scan by artificial intelligence 學生：張智鈞 指導教授：蔣依吾教授 國立中山大學資訊 程學系

肌少症 • 目前有發表論文定義肌少症的組織： • the European Working Group on Sarcopenia in Older People (EWGSOP) • the European Society for Clinical Nutrition and Metabolism Special Interest Groups (ESPEN-SIG) • the International Working Group on Sarcopenia (IWGS) • 共同定義症狀： • 低肌肉表現 • 低肌肉含量

電腦斷層掃描(CT) • Hounsfield units是描述放射密度的定量量表 Hounsfield units Tissue >1000 Bone, calcium, metal 100 to 600 Iodinated CT contrast 30 to 500 Punctate calcifications 60 to 100 Intracranial hemorrhage 35 Gray matter 25 White matter -29 to 150 Muscle, soft tissue 0 Water − 70 to − 30 Fat <− 1000 Air

深度學習 • 影像處理、音訊處理等非結構化資料 • 影像處理： • 影像分類（Image classification） • 影像偵測（Image detection） • 影像分割（Image segmentation） https: //medium. com/@prashant. brahmbhatt 32/the-yolo-object-detection-c 195994 b 53 aa

人 標記 • 經由專業醫生檢視CT圖並標記範圍 Psoas muscle Paraspinal muscle Lumborum quadratus

Ground Truth a b d c e

資料拆分 100%，1024張 60%，616張 Training data set 20%，204張 Validation data set 20%，204張 Testing data set

模型架構Unet • 編碼解碼網絡架構 (Encoder-Decoder network architecture) • 編碼器 • 解碼器 • 使用Skip connection結合抽象特徵

模型架構Unet • 每個CNN節點包含兩層 3× 3 具有相同邊界填充（Padding） https: //medium. com/machine-learning-world/convolutional-neural-networks-for-all-part-ii-b 4 cb 41 d 424 fd

The mean intersection over union (m. Io. U)

預測結果 AI model 16 Non- augmented avg 96% Augmented avg 97% Non- augmented avg 95. 1% Augmented avg 95. 8% 32 Cross validation 1 validation 2 validation 3 Average Filter size

預測結果 m. IOU Distribution CV 1, 204 test images 250 200 200 191 195 original filter size = 16 Count 150 original filter size = 32 augmentation filter size = 16 100 augmentation filter size = 32 50 0 0 2 0 0 0 1 4 4 11 8 0 50 60 70 m. IOU score (%) 80 90

預測結果 m. IOU Distribution CV 2, 204 test images 250 195 200 194 186 182 original filter size = 16 Count 150 original filter size = 32 augmentation filter size = 16 100 augmentation filter size = 32 50 0 0 0 2 0 3 9 10 16 17 0 50 60 70 m. IOU score (%) 80 90

預測結果 m. IOU Distribution CV 3, 204 test images 250 200 202 191 195 original filter size = 16 Count 150 original filter size = 32 augmentation filter size = 16 100 augmentation filter size = 32 50 0 0 0 0 2 0 1 4 2 11 8 0 50 60 70 m. IOU score (%) 80 90

預測結果較好的案例 miou=99% miou=97% miou=98% a b c d

預測結果較差的案例 miou=81% miou=88% miou=92% a b c d