Skip to main content
. 2020 Jun 23;296(3):584–593. doi: 10.1148/radiol.2020192091

Figure 4:

Knee radiographs overlaid with the heatmaps obtained by using the Grad-CAM method (27) (Appendix E1 [online]) show regions affecting the prediction of the multitask deep learning (DL) model trained with transfer learning. Colored regions show areas where multitask DL model trained with transfer learning focuses on decisions regarding the probability of total knee replacement in the patient within 9 years. Each column represents radiographs and heatmaps from a 75-year-old male control patient with Kellgren-Lawrence (KL) grade 0 (top and bottom left; body mass index [BMI], 26.0 kg/m2), a 76-year-old female control patient with KL grade 2 (top and bottom middle; BMI, 28.4 kg/m2), and a 70-year-old male patient with KL grade 2 (top and bottom middle; BMI, 27.6 kg/m2). Patients underwent total knee replacement and control patients did not undergo total knee replacement within 9 years.

Knee radiographs overlaid with the heatmaps obtained by using the Grad-CAM method (27) (Appendix E1 [online]) show regions affecting the prediction of the multitask deep learning (DL) model trained with transfer learning. Colored regions show areas where multitask DL model trained with transfer learning focuses on decisions regarding the probability of total knee replacement in the patient within 9 years. Each column represents radiographs and heatmaps from a 75-year-old male control patient with Kellgren-Lawrence (KL) grade 0 (top and bottom left; body mass index [BMI], 26.0 kg/m2), a 76-year-old female control patient with KL grade 2 (top and bottom middle; BMI, 28.4 kg/m2), and a 70-year-old male patient with KL grade 2 (top and bottom middle; BMI, 27.6 kg/m2). Patients underwent total knee replacement and control patients did not undergo total knee replacement within 9 years.