[笔记] Skull R-CNN: A CNN-based network for the skull fracture detection

文献信息

Skull R-CNN：基于 CNN 的网络用于颅骨骨折检测
Zhuo Kuang， 华中科技大
Xianbo Deng，华中科技大学同济医学院附属协和医院
MIDL 2020 会议论文
Impact Score: 1.32
H5-index: 0
Computer Science Conferences Ranking

Motivation

1. Previous studies for the automatic detection could not achieve the accuracy and robustness for clinical application;
   临床应用中以前方法准确性和鲁棒性不足
2. The sparse boxes cannot contain the small objects well and the candidate boxes are evenly distributed on the whole image, this will result in a lot of computational redundancy;
   Faster-RCNN中全图稀疏候选框无法包含小目标，较多计算冗余
3. CT scans, the skull fracture presents the following characteristics: (1) The fractures usually present as narrow slits; (2) The locations and the length of fractures are diverse; (3) A considerable percentage of the fractures have very small sizes.
   颅骨特征表现为狭缝、骨折多样和小尺寸

Contribution

1. Propose a novel method Skull R-CNN for more accurate skull fracture detection results;
2. A skeleton-based region proposal method is proposed to make candidate boxes more concentrated in key regions and reduced invalid boxes;
   基于骨架的区域建议方法
3. A novel full resolution feature network is constructed to obtain more precise features to make the model more snesetive to small objects;
   全分辨率特征网络
4. Skull R-CNN significantly reduces the false positives, and keeps a high sensitivity.
   提出Skull R-CNN减少FP，更高敏感性和更高准确率

Approach

1. Skull R-CNN
   First, multi-scale feature maps (from Full resolution feature network) and all candidate region boxes (from Skeleton-based region proposal);
   得到多尺度特征图和区域候选框
   Then, feature maps are adjusted to the same dimension with the help of the Roi-Align;
   Roi-Align调整得到相同维度
   Next, the classification module will screen out suitable boxes;
   分类网络对候选框打分和筛选
   Finally, the position of the boxes will be fined tune to better contain the fractures.

2. Full resolution feature network
   It is encoder-decoder structure with 3 times of downsampling and upsampling with the skip connection, and concatenation is applied here to merge the feature maps, in order to obtain the feature maps of full, quater, and eighth resolution.
   3次上采样-下采样的编码器-解码器网络，拼接合并特征，与FPN元素相加不同。网络输出原图的全分辨率、1/4和1/8。

3. Skeleton-based region proposal
   Skeletonization method to obtain the skull skeletons , the gridlines are defined as , anchors , the sizes of 8, 16, 32 and length-width ratio of 0.5, 1, 2 are applied to build the boxes around the anchors.
   锚位置用骨架方法引导，3种候选框尺寸和长宽比，不再需要RPN粗选。

4. Classification and regression
   The ROI-Align proposed in Mask R-CNN is to replace the ROI-Pooling. Then a classification network with fully connected layers scores all the boxes within 0 to 1 (binary classification), and the Non-Maximum Suppression (NMS) is adopted to get the boxes with the highest score in the local area.
   ROI-Align调整特征到相同维度，分类网络打分，非最大值抑制筛选得到最优框。
   The location offsets of these selected boxes are obtained, with a regression network.
   通过回归网络得到了这些选定方框的位置偏移量。

Experiment

1. labels of the candidate boxes
   label = 0 IOU 0.3; label=1 IOU0.5
   引入48个额外样本解决正负样本不平衡问题
   

2. 训练损失
   训练分两步
   Backbone and the classification module to train with Binary Cross Entropy (BCE) as the loss function.
   Then, the whole model is trained, and smoothL1 is set as the loss function for the regression module.
   The second step, to balance the two kinds of
   losses, the weighted loss .

3. 训练数据
   The CT scans from 45 head trauma patients were composed of 512 ×512 pixels with 1:25mm slice thickness. 872 slices were obtained for the experiment.
   CT切片512 ×512大小，总计872张
   数据转换时根据CT强度 (HU) 范围0-2550进行
   25 scans with 491 slices were selected as the training set, 10 scans with 208 slices as the validation set, 10 scans with 173 slices as the test set.
   训练集491，验证集208，测试集173

4. 测试指标
   PR曲线下AP值用来衡量检测性能。
   the detection results with the IOU greater than 0.5 were defined as the true object boxes.
   IOU大于0.5为正确目标框
   Detection time(s/slice)每个切片的检测时间

5. 比较方法
   Faster R-CNN + FPN
   Skull R-CNN + FPN
   Skull R-CNN
   3d black-Hat transform

Skull R-CNN improves the AP by an average of 0.075, as for the small objects, the improvements are around 0.11.

Conclusion

1. A novel Skull R-CNN was proposed for the automatic skull fracture detection and can accurately detect in a short time. It has less false positives and keep a high sensitivity;
2. A skeleton-based region proposal method was designed to replace the RPN;
3. A full resolution network was constructed to extract more precise features for the small objects detection;
4. The biggest constraints on performance is that some agittal, lambdoid and coronal sutures are also detected as fractures. Besides, sometimes the regression module can not give a precise offsets to the boxes;
5. Future works will focus on improving the ability of the proposed network to distinguish skull fractures from suture lines.

Reference