基于深度学习的腹内脂肪分割问题 - SJTU CS.PPT

Segmentation of Abdominal Adipose Tissues via Deep Learning(基于深度学习的腹内脂肪分割问题） Fei Jiang Abdominal Adipose Tissues Segmentation Task (a) MR image (b) segmentation Volumes of visceral adipose tissues (VAT) and subcutaneous adipose tissues (SAT) are better indicator of obesity-related diseases than BMI. Harm Background: AAT Segmentation Diagnosis Time-consuming 1.9 billiion overweight SliceOmatic: 4,000$ limitation：semi-automatic time：6-8min/slice Time-consuming, object-biase Not practical, expensive 39%（2014） Metabolic and cardiovascular diseases, Liver cancer… Segmentation of Abdominal Adipose Tissues Flowchart of Previous methods Previous methods Three steps: (1) abdomen mask; (2) separating SAT; (3) separating VAT Graph cut; Active contour; thresholding;… shortcomings Time-consuming re-training process for a new image Two steps for separating SAT and VAT, extra errors Our methods: AAT Segmentation Pixel-based classification Novelties: task-based algorithm. Once the classifier is trained, it could be directly applied to a new image. SAT and VAT separated by one step, no extra errors automatic, unsupervised feature representation learned by DNN Segmentation time: a few seconds/slice Accuracy: Dice Ratio: SAT: 0.94+/-0.02; VAT: 0.88+/-0.04 Results: AAT segmentation Tasks to do Software to automatic segmentation of abdominal adipose tissues Segmentation of 3-D abdominal adipose tissues Automatic detection of lung module Samples of lung and modules Image of lung modules Previous methods for detecting lung module Tasks to do 肺结节半自动检测 肺结节自动检测 肺结节自动检测软件 模板来自于 * 模板来自于 *