verseg

发表于 |

  1. challenge

    Large Scale Vertebrae Segmentation Challenge

    • task1:Vertebra Labelling，关键点检测
    • task2:Vertebra Segmentation，多类别分割

  2. data

    1. variation：数据affine轴不统一，尺寸不统一，扫描范围不统一，FOV区域不统一
    2. nii的两大解析工具：nibabel库load data的xyz顺序与axcode的顺序一致，e.g.[‘R’,’A’,’S’]的orientation会得到xyz的array，而sitk的读取刚好反过来，sitk的arr会是zyx。我们之前在将dicom写入nii时，会指定一个不为np.eye(4)的affine，就是为了transpose这三个轴。

  3. model

    1. team paper \

      • 三阶段：第一阶段，due to large variation FOV of the dataset，粗分割定位脊柱位置，第二阶段，higher resolution多类别关键点定位center，获得each located vertebra，第三阶段，二类分割for each located vertebra。

      • keywords：1. uniform voxel spacing：不要随意resize，todo: trilinear interp；2. on-the-fly data augmentation：using SimpleITK

      • 第一阶段：Spine Localization

        • Unet
        • regress the Gaussian heatmap of spinal centerline
        • L2-loss
        • uniform voxel spacing of 8mm

        • input shape：[64,64,128]，pad？

      • 第二阶段：Vertebrae Localization

        • SpatialConfiguration-Net
        • regress each located vertebra‘s heatmap in individual channel
        • resampling：bi/tricubic interpolation
        • norm：maxmin on the whole dataset
        • uniform voxel spacing of 2mm

        • input shape：[96,96,128]，z-axis random crop，xy-plane use ROI from stage1

      • 第三阶段：Vertebrae Segmentation

        • Unet
        • binary segment the mask of each vertebrae
        • sigmoid ce-loss
        • uniform voxel spacing of 1mm

        • input shape：[128,128,96]，crop origin image & heatmap image based on centroids

    2. reference paper\

      • 核心贡献：1.MIP：combines the information across reformations，3D to 2D，2. 基于判别器的训练机制：encodes local spine structure as an anatomical prior，加固椎块间类别&位置的spacial information

      • MIP：
        • localisation and identification rely on a large context
        • large receptive field
        • in full-body scans where spine is not spatially centred or is obstructed by the ribcage, such cases are handled with a pre-processing stage detecting the occluded spine

      • adversarial learning：

        • FCN用于分割
        • AE用于评估分割的好坏
        • do not ‘pre-train’ it (the AE)
        • loss：an anatomically-inspired supervision instead of the usual binary adversarial supervision (vanilla GAN)

      • 先说FCN——Btrfly Network

        • 建模成回归问题，每个关键点对应一个通道的高斯heatmap，背景channel为$1-max_i (y_i)$

        • 双输入双输出（sagittal & coronal）

        • 两个视角的feature map在网络深层做了融合，to learn their inter-dependency

        • Batch- normalisation is used after every convolution layer, along with 20% dropout in the fused layers of Btrfly

        • loss：l2 distance + weighted ce

          $\omega$ is the median frequency weighing map, boosting the learning of less frequent classes(ECB)

      • 再说判别器——Energy-based adversary for encoding prior

        • fully-convolutional：its predictions across voxels are independent of each other owing to the spatial invariance of convolutions

        • to impose the anatomical prior of the spine’s shape onto the Btrfly net

        • look at $\hat{Y}_{sag}$ and $\hat{Y}_{cor}$ as a 3D volume and employ a 3D AE with a receptive field covering a part of the spine

        • $\hat{Y}_{sag}$ consists of Gaussians：less informative than an image, avoid using max-pooling by resorting to average pooling

        • employ spatially dilated convolution kernels

        • mission of AE：predict the l2 distance of input and its reconstruction, it learns to discriminate by predicting a low E for real annotations, while G learns to generate annotations that would trick D

      • inference：

        • The values below a threshold (T) are ignored in order to remove noisy predictions
        • 用外积，$\hat{Y}=\hat{Y}_{sag}\otimes\hat{Y}_{cor}$
        • 每个channel的最大值作为centroids

      • experiments

        • 【IMPORTANT】10 MIPs are obtained from one 3D scan per view, each time randomly choosing half the slices of interest
        • 对于每个视角，每次随机抽取一半数目的slice用于计算MIP

  • similar local appearance：

  • strong spatial configuration：凡是涉及到椎块-wise的信息，从全局信息入手