A Novel Focal Ordinal Loss for Assessment of Knee Osteoarthritis Severity

Knee osteoarthritis (KOA) occurs mostly in the elderly and often causes physical disability and limitation. Early detection and intervention are particularly important in reducing the deterioration of knee disease. Currently, early detection of KOA is mainly evaluated by a Kellgren-Lawrence (KL) Classification with five class. KL is an ordered classification problem. The cross-entropy loss (CEL) and fixed penalty loss (FPL) are commonly used in KL system. Both CEL and FPL do not take into account the ease of classification between samples, which results in that the obtained model has deficiency. In this paper, a novel focal ordinal loss (FOL) is proposed by combining FPL and focal loss (FL) for KL. In the training algorithm based on FOL, the difficult or easy example is first identified according to the predicted probability toward to the true label at each epoch. The sample with a high predicted probability is considered as an easy sample. On the contrary, it is considered as a difficult sample. The weights for easy samples in the loss are then adjusted down in the next epoch, which results in that the training process can focus on difficult samples. The performance of FOL is validated on an X-ray image dataset from the Osteoarthritis Initiative (OAI) with several classical CNN models, such as VGG, Resnet, Densenet and Googlenet. The experimental results show that FOL achieves significant improvements in many performance measures, especially in mean square error (MSE). In addition, the experimental results on the augmented dataset and the Resnet with the convolutional block attention module (CBAM) also show similar improvements from FOL. It indicates that FOL is effective and superior to CEL and FPL in the numerous models and two data types (original and augmented) for KL grading.