Pretreat-MetsToBrain-Masks | A Large Open Access Dataset of Brain Metastasis 3D Segmentations on MRI with Clinical and Imaging Feature Information

DOI: 10.7937/6be1-r748 | Data Citation Required | Image Collection

Summary

Cohort Selection:

Patient images were collected from three sources: the Yale New Haven Health database (2013-2021), the Yale tumor board registry (2021), and the Yale gamma knife registry (2017-2021).

Inclusion/exclusion criteria:

Inclusion criteria included a pathologically proven diagnosis of brain metastasis and availability of a pretreatment scan with standard MRI sequences (T1 w, T1 post-gadolinium, T2w, and FLAIR). Exclusion criteria included lack of pretreatment scan or one of the standard MR sequences and significant motion artifact in any of the standard sequences.

Imaging acquisition/parameters:

Images were acquired using a 1.5T or 3T scanner. Most patients had T1 post-gadolinium sequences acquired using the MPRAGE protocol with a small subset having only spin echoes. Segmentations of core and necrotic tumor components were performed on T1 post-gadolinium sequences while whole tumor was segmented on FLAIR sequences.

Segmentation protocol:

All segmentations were performed manually on DICOM images in research PACS using a volumetric tool after transfer from clinical PACS.  All segmentations were approved by two neuroradiologists.

Image preprocessing:

The four standard sequences for each patient were exported from research PACS into NifTI format, co-registered to the SRl24 anatomical template, resampled to a uniform isotropic resolution (1 mm3), and skull stripped. Core, necrosis, and whole segmentation masks were exported individually from research PACS into NifTI format, combined into one mask while retaining their positions in native space, and registered to the SRl24 anatomical template before resampling to a uniform isotropic resolution. The sequence and segmentation NifTI files were manually checked and corrected by a neuroradiologist on the ITK-SNAP software.

Clinical and imaging data collection:

Demographic data, including sex, ethnicity, age at diagnosis, smoking pack-year history, and presence of extranodal metastasis were obtained using the electronic medical record (EMR). Survival was calculated by subtracting the date of diagnosis from the date of death or from the date of last EMR note for censored patients. Qualitative imaging features, including infratentorial involvement and intratumoral susceptibility in at least one lesion on SWI sequence, were obtained from visual assessment of images. Quantitative imaging features were extracted from the NifTI segmentation mask for each patient and included total enhancing tumor volume, total necrotic tumor volume, total peritumoral edema volume, ratio of necrotic to enhancing volume, ratio of peritumoral edema to enhancing volume, number of enhancing lesions, number of necrotic lesions, and number of lesions with peritumoral edema.  Finally, origin of metastasis was also obtained from the EMR from previous oncological and/or pathological reports. Dates of birth, diagnosis, death, and last note were all anonymized preserving duration between events. Clinical and imaging feature information are stored in an Excel file. 

Currently, many published brain metastasis segmentation algorithms have only been trained and validated on single institution datasets, leading to poor model generalizability. The availability of public datasets is important for algorithm generalizability and implementation into clinical practice. Our dataset is unique in its inclusion of many sub centimeter brain metastases, for which there are yet no robust algorithms. In addition, the dataset has many lesions with necrotic segmentations, which are historically not provided with dataset segmentations. We also include survival outcomes manually obtained from the EMR, allowing for future correlation of our imaging data with clinical outcomes. In addition, the heterogeneity of sources from which we obtained our patient data will allow for algorithms to be trained on a real-world hospital dataset that is not highly curated, thus enhancing model applicability to clinical practice. 

Data Access

Version 1: Updated 2023/12/19

Additional Resources for this Dataset

The following external resources have been made available by the data submitters.  These are not hosted or supported by TCIA, but may be useful to researchers utilizing this collection.

• CaPTk software was used to apply the BraTS Pre-processing Pipeline for image registration and brain extraction on these data https://cbica.github.io/CaPTk/preprocessing_brats.html

Citations & Data Usage Policy

Data Citation Required: Users must abide by the TCIA Data Usage Policy and Restrictions. Attribution must include the following citation, including the Digital Object Identifier: