An Automatic Knee Osteoarthritis Diagnosis Method Based on Deep Learning: Data from the Osteoarthritis Initiative

Figure 9

Comparisons of the attention maps of ResNet50 and the proposed method. The first row (Figure (a) to Figure (e)) shows the activated areas of the ResNet50 model generated by GradCAM. The second row (Figure (f) to Figure (j)) shows the attention weights of the visual transformer extracted through the attention flow technique. For each column, input X-ray images are the same. The proposed method succeeds in locating the narrow joint spaces on both sides of the knee. In addition, we detect sclerosis or bone spurs on the medial or lateral edge of the femur as in (f), (g), and (j).

(a)

(b)

(c)

(d)

(e)

(f)

(g)

(h)

(i)

(j)