Table 2.

Metrics to evaluate model performance.

Metric	Description
Intersection-over-Union (IoU)	The IoU metric reports on the overlap of output and ground truth segmentation masks. Higher overlap represents a better agreement between the model output and ground truth.
Precision and recall	These metrics are used to quantify the performance of instance segmentation or object detection. Precision is a measure for the specificity and describes which fraction of the detected objects are correctly detected/assigned. Recall, on the other hand, describes the sensitivity, i.e. how many objects out of all objects in the dataset were detected.
(mean) average precision ((m)AP)	This metric is used to evaluate model performance in object detection and classification tasks. It describes the models’ ability to detect objects of individual classes (AP) or all classes (mAP) present in the dataset. To obtain the average precision, precision and recall values for the individual object classes are calculated at different detection thresholds. mAP is calculated by averaging all single-class AP values.
Structural similarity (SSIM)	The SSIM value quantifies how similar two images are with respect to pixel intensities and intensity variations. As it is calculated locally using a defined windows size, it provides a similarity map that allows to identify regions of high or low similarity.
Peak-signal-to-noise ratio (PSNR)	The PSNR metric compares the signal to noise ratio of images with lower signal-to-noise to the high SNR counterpart based on the pixel-wise mean squared error. It is often used to compare the results of image compression algorithms, but can also be applied to evaluate model performance on paired test data.