Automatic-Knee-Contour-Detection-Using-U-net-Segmentation

U-net based BioMedical Image Segmentation approach for Knee X-rays to automate contour extraction on femur bone When working at AlgoSurg I came across this interesting problem of automating contour extraction of knee-bones, these contours help drive surgeon's decision about the surgery, manually annotating takes lot of efforts and time for a fairly repetitive task thus neural networks come to rescue

My solution is based on main U-net architecture but there are some fine changes that I implemented by getting insprired from the papers mentioned in references

• I have used same filters model thus keeping filters=128 throughout the U-net
• I have used Average Pooling instead of maxpooling

Dataset is not included in the github as it is proprietary Thanks to AlgoSurg for letting me publish this code

I first tried to solve this problem using landmark detection but I soon realised that the problem is more similar to edge detection and therefore I used segmentation type approach and optimized Binary Cross Entropy Loss I tried many different architectures such as Dilated Residual U-net, Stack U-net

References

Heatmap Regression 2016
Heatmap Regression 2019
Christian Payer et al. Github
2D Dilated Residual U-net
Stack U-net

Output

Legends

• Middle one is ground truth hand annotated
• Right most is the predicted image
• Red points are predicted points extracted and downsampled from predicted image
• Blue Points are hand annotated points

1.

2.

3.

4.

5.

6.