Fast PET-CT lesion detection (fastPET-LD)

Background

PET-CT has become the tool of choice for oncologic imaging, providing unrivalled detection and follow up of primary tumors and metastases. A radiotracer, that is a molecule with a positron emitting part, is injected to the body and accumulates at targeted sites. For example, 18FDG, a modified glucose with a radionuclide fluorine 18, is commonly used to detect increased glucose uptake in cancer cells. The photon pairs generated by the annihilation of the radiotracer’s protons are collected at diametrically opposed detector surrounding the body. A tomographic reconstruction is performed from the collected photon pairs to generate axial 2-D maps of the tracer distribution within the whole body. PET and CT are generally integrated into a single machine to allow for CT based attenuation correction of PET images and obtain accurate geometrical fusion between the 2 modalities. Detecting “hot spots” and discriminating real lesions from false positive is usually performed with the help of the overlaid CT scan.

An important bottleneck of PET imaging is the long exposure time required to collect enough photons at the detectors to reach an acceptable SNR in the reconstructed images. Current commercial systems require 75-150 seconds of exposure per-bed-position (pbp) to provide clinical grade images.

Significantly reducing the exposure time, and from there, the total exam time is a must to improve poor throughput, increase patient experience, reduce motion artifacts, and improve care providers profitability. However, the resulting lower quality images may affect diagnostic ability.

Challenge

In this challenge, we provide 2 training datasets of 68 cases each: the first one was acquired at Sheba medical center (Israel) nuclear medicine department with a very-short exposure of 30s pbp, while the second is the same data followed by a denoising step implemented by a fully convolutional Dnn architecture trained under perceptual loss [1,2]. The purpose of this challenge is the detection of “hot spots”, that is locations that have an elevated standard uptake value (SUV) and potential clinical significance. Corresponding CT scans are also provided. The ground truth, common to both datasets, was generated by Dr. Liran Domachevsky, chair of  nuclear medicine  at Sheba medical center. It consists of a 3-D segmentation map of the hot spots as well as an Excel file containing the position and size of a 3D cuboid bounding box for each hot spot.

Out of 68 cases, 55 with  ground truth, for training,  while 13 cases, the test set, are provided without ground truth.

More details in participation section.


Data usage agreement

Participants cannot share the data, cannot use it for any commercial purpose. If this dataset or part of it is used in a published
paper (as well as test set evaluation results from the leaderboard) please cite the following paper:

@inproceedings{green2019feature,
  title={Feature Aggregation in Perceptual Loss for Ultra Low-Dose (ULD) CT Denoising},
  author={Green, Michael and Marom, Edith M and Konen, Eli and Kiryati, Nahum and Mayer, Arnaldo},
  booktitle={2019 IEEE 16th International Symposium on Biomedical Imaging (ISBI 2019)},
  pages={1635--1638},
  year={2019},
  organization={IEEE}
}

Contacts

If you have any questions, please contact:


References

[1]  Zhang K, Zuo W, Chen Y, Meng D, Zhang L. Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising. IEEE transactions on image processing. 2017 Feb 1;26(7):3142-55.

[2]  Green M, Marom EM, Konen E, Kiryati N, Mayer A. Feature Aggregation in Perceptual Loss for Ultra Low-Dose (ULD) CT Denoising. In2019 IEEE 16th International Symposium on Biomedical Imaging (ISBI 2019) 2019 Apr 8 (pp. 1635-1638). IEEE.