Researchers at the University of Central Florida (UCF)’s Center for Research in Computer Vision have created an artificial intelligence (AI) algorithm that can detect specks of lung cancer in CT scans with 95 percent accuracy.
“I believe this will have a very big impact,” Ulas Bagci, assistant professor of engineering and imaging scientist at the University of Central Florida, said in a UCF statement. “Lung cancer is the number one cancer killer in the United States and if detected in late stages, the survival rate is only 17 percent. By finding ways to help identify earlier, I think we can help increase survival rates.”
Bagci and colleagues said humans can detect these cancers with only 65 percent accuracy.
The researchers wrote, even with computer aided detection, automatic detection of lung nodules is challenging because of the relative small size of the cancers and variations in texture, shape and position of nodules.
“Detection of tiny/small objects has remained a very challenging task in computer vision, which so far has only been solved using computationally expensive multi-stage frameworks,” the researchers wrote.
Bagci developed S4ND, an algorithm to detect lung nodules, with UCF colleague Naji Khosravan. Instead of multi-stage frameworks, S4ND uses a single feed forward pass of a single network for detection. S4ND, the researchers said, does not require any post-processing or user guidance to refine detection results.
In their initial experiment to test the efficacy of S4ND, Bagci and Khosravan compared their algorithm with the current leading method for lung nodule detection—3D convolutional neural network (3D CNN).
Utilizing 888 CT scans that were publicly available from the Lung Nodule Analysis, they found S4ND outperformed 3D CNN in efficiency and accuracy. S4ND attained 95.2 percent accuracy, compared to 3D CNN’s 94.6 percent.
“We present a fundamental solution to address the major challenges of current region proposal-based lung nodule detection methods: candidate detection and feature resampling stages,” the authors concluded. “We experimentally validate the proposed network’s performance both in terms of accuracy (high sensitivity/specificity) and efficiency (less number of parameters and speed) on the publicly available LUNA data set, with extensive comparison with the natural object detector networks as well as the state-of-the-art lung nodule detection methods. A promising future direction will be to combine diagnosis stage with the detection.”
As a senior news writer for TriMed, Subrata covers cardiology, clinical innovation and healthcare business. She has a master’s degree in communication management and 12 years of experience in journalism and public relations.