Fig. 3
From: The impact of pre-processing techniques on deep learning breast image segmentation
Original image of a patient’s breast from the C6BIS-DDSM dataset, showing: (a) post-processed image using the best-performing model, DS with input size 256 x 256 and P1-P99 of intensity normalization, with the label lesion (shown in orange) and the model’s predictions (shown in blue), overlapped on the image; (b) post-processed image from the least effective model, DNS with input size 1024 x 1792 and Min-Max of intensity normalization, with the same ground truth label (orange) and the corresponding model predictions (blue), also overlapped.
