Deep learning model shows potential for lung tumor detection on CT scans
by
Gus Iversen, Editor in Chief | January 24, 2025
Artificial Intelligence
CT
X-Ray
A deep learning model has demonstrated promising results in detecting and segmenting lung tumors on CT scans, according to a study published in Radiology.
The study highlights the challenges of accurately identifying and delineating lung tumors, a process that is essential for tracking disease progression, assessing treatment response, and planning radiation therapy. Currently, this task is performed manually by clinicians, which can be time-consuming and subject to variability.
While AI models have been explored for this purpose, earlier studies were limited by small data sets and manual inputs, often focusing on single tumors rather than the complexities of multi-tumor cases. To address these challenges, researchers used a large-scale data set of pre-radiation treatment CT simulation scans and their corresponding 3D tumor segmentations.
The study, led by Dr. Mehr Kashyap, a resident physician at Stanford University School of Medicine, involved training an ensemble 3D U-Net deep learning model on 1,504 CT scans containing 1,828 segmented lung tumors. The model was tested on an additional 150 CT scans, with its performance compared to that of human physicians.
The model achieved a sensitivity of 92% and specificity of 82% in detecting lung tumors, with a median Dice similarity coefficient (DSC) of 0.77 when compared to physician-delineated segmentations. By contrast, the median DSC for physician-to-physician segmentations was 0.80. Notably, the model required less time to complete segmentations than physicians.
Dr. Kashyap emphasized the advantage of the 3D U-Net architecture, noting its ability to capture interslice information. "By capturing rich interslice information, our 3D model is theoretically capable of identifying smaller lesions that 2D models may be unable to distinguish from structures such as blood vessels and airways," he said.
However, the study noted a limitation: the model's tendency to underestimate tumor volume, particularly for larger tumors. Dr. Kashyap cautioned that the model should be integrated into workflows where physicians can supervise and correct potential errors.
Future research will explore the model’s ability to estimate total tumor burden, evaluate treatment response, and predict clinical outcomes when combined with other prognostic tools.
|
|
You Must Be Logged In To Post A Comment
|