GBCU is the first public dataset for Gallbladder Cancer identification from Ultrasound images. GBCU contains a total of 1255 (432 normal, 558 benign, and 265 malignant) annotated abdominal Ultrasound images collected from 218 patients. Of the 218 patients, 71, 100, and 47 were from the normal, benign, and malignant classes, respectively. The sizes of the training and testing sets are 1133 and 122, respectively. To ensure generalization to unseen patients, all images of any particular patient were either in the train or the test split. The number of normal, benign, and malignant samples in the train and test set is 401, 509, 223, and 31, 49, and 42, respectively. The width of the images is between 801 and 1556 pixels, and the height is between 564 and 947 pixels due to the cropping of patient-related information. We acquired data samples from patients referred to PGIMER, Chandigarh (a tertiary care referral hospital in Northern India) for abdominal ultrasound examinations of suspected Gallbladder pathologies. The study was approved by the Ethics Committee of PGIMER, Chandigarh. We obtained informed written consent from the patients at the time of recruitment, and protect their privacy by fully anonymizing the data. Grayscale B-mode static images, including both sagittal and axial sections, were recorded by radiologists for each patient using a Logiq S8 machine.
Each image is labeled as one of the three classes - normal, benign, or malignant. The ground-truth labels were biopsy-proven to assert the correctness. Additionally, in each image, expert radiologists have drawn an axis-aligned bounding box spanning the entire GB and adjacent liver parenchyma to annotate the region of interest. Additionally, bounding-box annotation for abnormal pathologies (e.g. stone, benign mural thickening, or malignancy) are also provided. The GBCU dataset is suitable for both image classification and object detection tasks. Apart from the Gallbladder Cancer, the dataset can also be used for detection of several other pathologies.
To obtain the dataset, please fill and sign this License Agreement, and send it to Dr. Pankaj Gupta and Dr. Chetan Arora. After duly verifying the License Agreement, we will mail the dataset download link to you. Note: We will only accept requests from permanent employees/ faculty of the requesting institute. We will ignore all requests from students. Student Researchers must ask their Supervisor/ Head of the Department to fill out and send the agreement to us.
@InProceedings{basu2022surpassing,
title={Surpassing the Human Accuracy: Gallbladder Cancer Detection from USG with Curriculum Learning},
author={Basu, Soumen and Gupta, Mayank and Rana, Pratyaksha and Gupta, Pankaj and Arora, Chetan},
booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
pages = {20854-20864},
year={2022}
}