Purpose: Radiomic features are promising quantitative parameters that can be extracted from medical images and employed to build machine learning predictive models. However, generalizability is a key concern, encouraging the use of public image datasets. We performed a quality assessment of the PROSTATEx training dataset and provide publicly available lesion, whole-gland, and zonal anatomy segmentation masks.
Method: Two radiology residents and two experienced board-certified radiologists reviewed the 204 prostate MRI scans (330 lesions) included in the training dataset. The quality of provided lesion coordinate was scored using the following scale: 0 = perfectly centered, 1 = within lesion, 2 = within the prostate without lesion, 3 = outside the prostate. All clearly detectable lesions were segmented individually slice-by-slice on T2-weighted and apparent diffusion coefficient images. With the same methodology, volumes of interest including the whole gland, transition, and peripheral zones were annotated.
Results: Of the 330 available lesion identifiers, 3 were duplicates (1%). From the remaining, 218 received score = 0, 74 score = 1, 31 score = 2 and 4 score = 3. Overall, 299 lesions were verified and segmented. Independently of lesion coordinate score and other issues (e.g., lesion coordinates falling outside DICOM images, artifacts etc.), the whole prostate gland and zonal anatomy were also manually annotated for all cases.
Conclusion: While several issues were encountered evaluating the original PROSTATEx dataset, the improved quality and availability of lesion, whole-gland and zonal segmentations will increase its potential utility as a common benchmark in prostate MRI radiomics.
Keywords: Machine learning; Prostate MRI; Public imaging dataset; Radiomics; Segmentation.
Copyright © 2021 Elsevier B.V. All rights reserved.