Proud to be part of the lung cancer nodule detection challenge winning team with my colleagues Jean-Armand (Cognitive Systems Europe team & Team Tech leader), Maxime (GBS Interactive working at Montpellier Cognitive Systems Lab), Thibaud (FTSS Cognitive Systems France) and Illyas (trainee@Montpellier Cognitive Systems Lab) and our partner Quantacell https://www.quantacell.com/ .
The objective of the challenge was to "distinguish 3D scanner exams with lung cancer nodules greater than 100 "
Here are links (in French) to the challenge description https://jfr.radiologie.fr/les-jfr/villages-et-forum/forum-intelligence-artificielle and to a medical imaging journal article discussing results https://docteurimago.fr/actualites/socioprofessionnel/le-data-challenge-2019-recompense-six-laureats/ .
About the challenge and the solution
The lung cancer nodules are usually diagnosed late because of the absence of characteristic symptoms; which makes lung cancer the leading cause of mortality in France
These data were transmitted to us via 3 successive datasets of approximately equal size (about 170 pathological images and 170 normal images each). A first test set was given on September 11, 2019 served as a training set; a second set on October 11, 2019 served as validation. Finally, the third dataset, final test on 344 images with only one hour to process.
It is a difficult detection problem because of the respiratory and circulatory systems network, leading to potential confusion.
Our approach was to implement a pre-processing function that handles original images and annotations, isolating the lungs from the rest of the body using Hounsfield Units (HU. This detection allowed t us to automatically reject artifacts detected outside the lungs.
Then followed a cancer nodule detection model, based on the « Retina U-Net 3D » architecture https://www.nature.com/articles/s41591-019-0447-x , that accepted the preprocessed image input and predict the detected nodules by generating the coordinates of a rectangular parallelepipedal box, but also intensity attributes and segmentation results.
Finally, a classification model, which from the list of attributes predicted for a patient, will classify it as pathological (having at least one nodule greater than 100 mm3) or normal.
We used an orchestrator aggregator architecture on 3 IBM Power AC922 servers with Nvidia V100 32 Go graphic cards (GPU – Graphics Processing Unit) (respectively 4, 4 et 6, for a total of 14 GPUs) https://www.ibm.com/us-en/marketplace/power-systems-ac922 .
That processing power together with the capacity to handle large models allowed to process successfully the test images in the required time with the best result of the ten teams in the challenge.
#GlobalAIandDataScience#GlobalDataScience