April 26, 2019
AI & Industry 4.0
Ciklum News

Ciklum Research Engineer Develops Machine Learning Model to Segment Organs for Radiotherapy Planning

Ciklum Research Engineer Develops Machine Learning Model to Segment Organs for Radiotherapy Planning

Maksym Manko, a Ciklum research engineer, has developed a machine learning model that segments organs in the chest cavity (esophagus, heart, trachea, aorta) with high accuracy. While subjecting patients to radiotherapy, it is important to delineate organs precisely so the radiation dose is not directed at OAR, healthy Organs at Risk, near a tumour. The model enables more efficient planning of radiation therapy for cancer treatment.

Traditionally, the delineation of OAR is manual, which is both time-consuming and subjective. For some organs such as the esophagus, segmentation comes with many challenges e.g. organ shape, size and position may significantly vary between different patients, contours in CT images have low contrast and can even be invisible.

Maksym used the dataset provided by ISBI, consisting of 60 CT scans (40 CT scans for network training, 20 CT scans for evaluation). He gave preference to 3D segmentation, which takes into account spatial information and predicts object boundaries faster than a radiologist, providing them with more bed-side time with patients. After a thorough analysis of the publications on the subject, Maxym chose the U-Net architecture, which is considered to be the state-of-the-art neural network architecture for biomedical image segmentation tasks.

machne learing model for organs at risk segmentation

Boundaries of the segmented organs on CT scans from the test set

To assess the accuracy of the system, dice scores and the Hausdorff distance metrics were used. The following metric scores were obtained on the test split: dice score – 0.75 (esophagus), 0.94 (heart), 0.9 (trachea), 0.91 (aorta); Hausdorff distance – 0.67 (esophagus), 0.18 (heart), 0.3 (trachea), 0.32 (aorta).

A pre-trained deep learning model can segment organs automatically in less than 20-25 seconds per CT scan and work autonomously in the future. 

Maksym Manko
I’m honored that my 'Segmentation of Organs at Risk in Chest Cavity Using a 3D Deep Neural Network' paper for SPSympo-2019 has been accepted and is preliminarily assigned to be presented at a symposium session titled ‘Artificial intelligence and Machine Learning’. I continue to work on improving the model. The main challenge at the moment is to improve the accuracy of segmentation of the esophagus. It’s exciting and rewarding to work on something that will save lives
Maksym Manko, research engineer at Ciklum