the skull fracture detection MIDL 2020 Z H U O K UA N G 1 ; X I A N - - PowerPoint PPT Presentation

Z H U O K UA N G 1 ; X I A N B O D E N G 2 ; L I Y U 1 ; H A N G Z H A N G 2 ; X I A N L I N 1 ; H U I M A 2

1 H U A Z H O N G U N I V E R S I T Y O F S C I E N C E A N D T E C H N O L O G Y , C H I N A 2 U N I O N H O S P I T A L A F F I L I A T E D W I T H T O N G J I M E D I C A L C O L L E G E O F H U A Z H O N G U N I V E R S I T Y O F S C I E N C E A N D T E C H N O L O G Y , C H I N A

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

MIDL 2020

Content

Background Method Experiment results

Background

Background

Figu

• Figure. 1.

. The skull fractures annotated by the radiologist. The blue boxes are the ground truth annotated by the radiologists, which contain the fractures.

Background

Fig

• Figure. 2.
• 2. The distribution of the width and length of the object boxes.

The fractures usually present as narrow slits; The locations and the length of fractures are diverse; A considerable percentage of the fractures have very small sizes;

Method

Skull R-CNN

Full resolution feature network Roi Align Skeleton-based Region Proposal Proposal Layer Classification Regression Feature Maps Coordinates

64*512*512 64*256*256 64*128*128

Fig

• Figure. 3.
• 3. The architecture of the Skull R-CNN

Skeleton-based region proposal

A B C D

Based on the feature map with low resolution Based on the origin CT image

Fig

• Figure. 4.
• 4. Le

Left: Region proposal network(RPN)[1]; Righ Right: Skeleton-based region proposal

The candidate boxes are much less than RPN, while keeping enough boxes containing fractures. Compared to RPN, there is no need to be trained and it just costs small amount of computation.

• 1. Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real- time object detection with region proposal networks. In

Advances in neural information processing systems, pages 91–99, 2015.

Full resolution feature network

64 64*5 *512 12*51 512 64 64*2 *256 56*25 256 128*256 128*256*256 *256 12 128* 8*12 128*1 *128 28 25 256* 6*12 128*1 *128 28 256*64* 256*64*64 64 256 256*12 *128*12 *128 256 256*12 *128*12 *128 128 128*25 *256*25 *256 128 128*25 *256*25 *256 64* 64*512 512*512 512 64*512* 64*512*512 512 64* 64*512 512*512 512 64* 64*256 256*256 256 64* 64*128 128*128 128

Upsampling(by2) Max Max-pooling(by2) Conv(3*3+BN+ReLu)

64* 64*256 256*256 256 128 128*12 *128*12 *128 256 256*64 *64*64 64 1*5 1*512* 12*512 12

Inut image Prediction Prediction Prediction Conv(1*1+BN+ReLu)

Fig

• Figure. 5.
• 5. The structure of the full resolution feature network.

The output feature maps have higher resolutions than the FPN[2], and have more accurate local information; Compared to FPN[2], element-wise addition is replaced by the concatenation to softly merge the feature maps.

2.Lin T Y , Dollár, Piotr, Girshick R , et al. Feature Pyramid Networks for Object Detection[J]. 2016.

Experiment results

Objective indices

Methods AP(×0.01) Detection time(s\slice) val test val(<16*16) test(<16*16) val test Faster R-CNN + FPN 55.7 54.2 59.4 49.3 0.088 0.087 Skull R-CNN + FPN 62.6 57.9 64.7 58.6 0.058 0.058 Skull R-CNN 65.1 60.0 67.3 63.3 0.035 0.036

Fig

• Figure. 6.
• 6. The PR curves on the test set. Le

Left: Faster R-CNN+FPN; Righ Right: Skull R-CNN Tab

• able. 1.

. The performance of the models.

Subjective results

Fig

• Figure. 7.
• 7. The detection results of the Skull R-CNN. The images in the second row are the partial magnifications of images in the fist row. In which, the

green boxes are TP predictions, the red boxes are FP predictions, and the blue boxes are the FN predictions

Thank you for listening !