KL-Grade-classification

• This project is conducted as a part of Healthhub AI Research Center.

Dual Input Model

• Medial and lateral images are convoluted separately, and concatenated afterwards. Dense layer makes predictions afterwards
• Histogram Equalization & Normalization.
• Multitasking. We trained the network to make a second prediction: rather the X ray captures the left or right knee. The model converged very quickly. Loss weight 0.75:0.35 for KL grade and side, respectively, yielded the highest accuracy. Mean accuracy 70% (5-fold validation)
• Result Summary: Mean accuracy 80% (5-fold cross validation). Near perfect accuracy for Class 2.

Autoencoder

Adversarial Autoencoder

• Adversarial autoencoder that is trained to follow Normal distribution (0,1).
• Autoencoder model reconstructs original image. Kl grader predicts KL grade of the input image using latent vector.
• 50 % accuracy