Our model predicting the target location in supine has reached a prediction accuracy of 14.2 +/-1.4 mm (mean +/- SD) for the median error and 16.3 +/-1.5 mm for the mean error (Figure 2). Accuracy was calculated as the distance between the predicted and actual target location in the supine position.

To understand how the ML model determines supine target position, we explored the importance of individual predicators by measuring their influence on the developed model. The most important predictor of the target location in the supine position is the displacement of the nipple from prone to supine, especially along the transverse axis. Other top predictors are (1) the angle between the vector from the nipple point in prone to its projection on the chest wall in prone and the vector from the nipple point in supine to its projection on the chest wall in prone and (2) the distance from nipple to sternum in prone.

There are several sources of error that can decrease accuracy in this prediction model. One type is the error that introduces variability in the dataset and cannot be explained by our prediction model. This type of error includes the error in the prone-to-supine image registration due to patient positioning differences, the respiratory motion error, and the error due to variations in the compressive deformation of the breast by the chest coil used to obtain the supine MRI images. The second type of error is the bias in the prediction model itself. This can be measured by examining correlations between predictors and individual target error. From this analysis we learned that our prediction model becomes less accurate when the target is far from the nipple. One option to improve our model would be to collect additional patient data, specifically for targets far from the nipple.