Knee Cartilage Segmentation using Improved U-Net

joint stability is a complex problem and requires detailed anatomic parametric study for knowing the associated breakdowns of knee cartilage. Osteoarthritis is one of the main disorders, which disrupt the normal bio-mechanics and stability of the patello-femoral joint and for diagnosing osteoarthritis radiologists needs a lot of time to diagnose it. An improved network called PSU-Net is proposed for the automatic segmentation of femoral, tibia, and patella cartilage in knee MR images. The model utilizes a Squeeze and Excitation block with residual connection for effective feature learning that helps in learning imbalance anatomical structure between background, bone areas and cartilage. The severity of knee cartilage is measured through the Kellgren and Lawrence (KL) grading system by radiologists. Also, updated weighted loss function is used during training to optimize the model and improve cartilage segmentation. Results demonstrate that PSU-Net can accurately and quickly identify cartilages compared to the traditional procedures, aiding in the treatment planning in a very short amount of time. Future work will involve the use of augmentation methods and also use this architecture as a generator model for generative adversarial network to improve performance further. The utility of this work will help in analyzing the anatomy of the human knee by the radiologists in short amount of time that may prove helpful to standardize and automate patello-femoral measurements in diverse patient populations.