Revert "Updated readme for DFCI examples"

README.md

@@ -62,11 +62,12 @@ The BraTS 2020 challenge dataset is a retrospective collection of 2,640 brain gl
 
 **Task**
 
-Precise organ segmentation using computed tomography (CT) images is an important step for medical image analysis and treatment planning. Pancreas segmentation involves important challenges due to the small volume and irregular shapes of the areas of interest. Our goal is to perform federated evaluation across two different sites using MedPerf for the task of pancreas segmentation, to test the generalizability of a model trained on only one of these sites
+Precise organ segmentation using computed tomography (CT) images is an important step for medical image analysis and treatment planning. Pancreas Segmentation involves immense challenge due to the small volume and irregular shapes. Our goal is to perform federated evaluation across different sites using MedPerf for the task of pancreas segmentation.
 
 **Data**
 
-We utilized two separate datasets for the pilot experiment. The first of which is the Multi-Atlas Labeling Beyond the Cranial Vault (BTCV) dataset, which is publicly available through synapse platform. Abdominal CT images from 50 metastatic liver cancer patients and the postoperative ventral hernia patients were acquired at the Vanderbilt University Medical Center. The abdominal CT images were registered using NiftyReg.  In addition to the BTCV dataset, we also included another publicly available dataset from TCIA (The Cancer Imaging Archives) at the url (https://wiki.cancerimagingarchive.net/display/Public/Pancreas-CT); the National Institute of Health Clinical Center curated this dataset with 80 abdominal scans, from 53 male and 27 female subjects. Of which 17 patients had known kidney donations that confirmed healthy abdominal regions, and the remaining patients were selected after examination confirmed that the patients had neither pancreatic lesions nor any other significant abdominal abnormalities. Ourl visual inspection of both datasets confirmed that they were appropriate and complementary candidates for the pancreas segmentation task. Out of the 130 total cases, we used 10 subjects for inference, 5 from each dataset. We arbitrarily chose patients of IDs 1 to 5 from each one.
+We utilized two separate datasets for the pilot experiment. The first of which is the Multi-Atlas Labeling Beyond the Cranial Vault (BTCV) dataset. This dataset is publicly available through synapse platform. Abdominal CT images from the metastatic liver cancer patients and the postoperative ventral hernia patients were acquired at the Vanderbilt University Medical Center. Voxel size for images was 0.6 to 0.9 mm in the anterior-posterior (AP) and left-right (LR) axis and 1.5 to 7.0 mm in the inferior-superior (IS) axis were the image acquisition parameters. Abdominal CT images were registered using NiftyReg. A total of 3719 images were obtained from 40 subjects for the task. 3719 images were randomly distributed into 2916 images for training, and 803 images for testing. The data distribution was done in a subject-wise manner to avoid data leakage between the training and the testing dataset. Due to the inconsistency in the image orientation, all the images were re-oriented to a standard orientation for further analysis.
+In addition to the BTCV dataset, we also included another publicly available dataset from The Cancer Imaging Archives (TCIA). The National Institute of Health Clinical Center curated the dataset with 82 abdominal scans, from 53 male and 27 female subjects. Of which 17 patients had known kidney donations that confirmed healthy abdominal regions, and the remaining patients were selected after examination confirmed that the patients had neither pancreatic lesions nor any other significant abdominal abnormalities. These scans varied between 1.5 - 2.5 mm, with 512 x 512 pixel resolution, generating 18782 individual scans.
 
 **Code**