Manual label
Thank you very much for sharing your data. I have two questions regarding it. First, the number of validation sets mentioned in the figures of your article is 1564, but the actual number downloaded is 1304. Could you please explain this discrepancy? Second, you mentioned in your article that an automatic annotation algorithm was used to label the data and that 1000 reports were manually annotated for training and validation. However, in the open-source data, it is not specified which data was manually annotated and which was automatically annotated. Could you clarify this?
Hi @Laihaoran , 1564 is number of CT scans, where 1304 is number of patients. Same patient might have more than 1 CT scan. I believe this is written in the text and in the figure 2.
The automatic annotation algorithm is trained on another set of reports from the same hospital randomly. They might and might not be in the CT-RATE dataset (most of them are in the dataset though). We additionally provide them in the GitHub page: https://github.com/ibrahimethemhamamci/CT-CLIP/tree/main/text_classifier/data. In the CT-RATE, all labels are extracted with the model. We also manually checked the extracted labels and they seem to be quite good.
Thank you for your reply. The validation set I downloaded consists of 1,304 patients. If calculated by the number of CT scans, there are a total of 3,039 CT volumes; if we do not consider different reconstruction parameters and only consider different CT experiments, the result is 1,638, which does not match the 1,564 CT volumes described in the article.
This gives 1564 CT scans (which does not include different reconstruction and reported in the figure, number of volumes contains reconstruction which is 3039 as you mention):
import pandas as pd
df = pd.read_csv("dataset_multi_abnormality_labels_valid_predicted_labels.csv")
names = {}
for index, row in df.iterrows():
name = row["VolumeName"]
name_final = name.split("_")[1] + "_" + name.split("_")[2]
print(name_final)
names[name_final]=0
print(len(names.keys()))
The names should be in the format valid_patientnumber_scannumber_reconstructionnumber.npz. So we need to count unique patientnumber_scannumber to find number of scans. Let me know if you see any problem with this.