Data Description¶
Task:
Detection of abnormalities: Can you detect the abnormal regions in a gastroscopic image?
Dataset:
We totally collect about 800 gastroscopic images from 137 volunteers. Three senior experts are invited to annotate the lesion/abnormal region independently, and we average them to generate the pixel-level ground truth.
The data is split into a training and test set (see below).
Training Data¶
Dataset | Image Num | Volunteer Num | Resolution | Ground Truth | ||
Training Data |
Abnormal | Normal | Lesion | Health | 489*409 | Pixel-level |
260 | 205 | 98 | 71 |
Test Data¶
Dataset | Image Num | Volunteer Num | Resolution | Ground Truth | ||
Test Data |
Abnormal | Normal | Lesion | Health | 489*409 | Pixel-level |
129 | 104 | 83 | 62 |
Reference Standard¶
For evaluation, we test the performance based on the image-level predictions. For positive images, one image can be considered as true positive if at least 40% of the truly abnormal pixels are detected, otherwise it will be considered as false negative. For negative images, one image can be considered as true negative only when no abnormal pixel is detected, otherwise it will be considered as false positive. Let
- TP be the number of true positives,
- FP be the number of false positives,
- TN be the number of true negatives,
- FN be the number of false negatives,
TPR = TP / (TP + FN), FPR = FP / (FP + TN), We choice different thresholds and compute the TPR and FPR accordingly to generate the Receiver Operating Characteristic (ROC) curve. Furthermore, the Area Under Curve (AUC) is used for statistical evaluation.