Machine learning aids in detecting lung contour, reducing radiologist workload

Radiation therapy is an integral part of many cancer treatments. Ideally, doses are focused on the observable tumor while leaving surrounding organs unaffected, but determining the figuration of tumors and organs-at-risk is done manually—a time consuming and, at times, imprecise task for radiologists.

A team of Chinese researchers developed a machine learning technique—closed polygonal line and backpropagation neural network model (CPL-BNNM)—for accurately detecting smooth lung contours in 3D-CT scans that is more efficient than manually determining such information and superior to currently used algorithms.

“The important information for organ diseases can be quantitatively provided by the clinical images, while quantification is often manually implemented in some clinics,” wrote Tao Peng, with the School of Computer Science & Technology at Soochow University. “In order to speed up the manual task and reduce workload, combining computer-aided diagnosis with automatic detection method is becoming a research hotspot.”

The study, published online Feb. 15 in the Journal of Digital Imaging, used the devised method to analyze data from the private clinical high-resolution lung dataset and 100 images from the public Lung Image Database Consortium and Image Database Resource Initiative (LIDC-IDRI).

In analysis done using the private data-set the Dice coefficient reached up to 0.95, with low global error. And in the LIDC-IDRI dataset the method achieved “superior segmentation performance” with an average Dice score of 0.83, according to Peng et al.

“With the proposed CPL-BNNM algorithm, the computational complexity of the contour extraction and the workload of radiologists can be reduced,” wrote Peng et al. “The quantitative and qualitative experimental results show that our proposed semi-automatic detection method has better extraction accuracy for high-resolution lung datasets obtained by 3D-CT scans.”