A new artificial intelligence model has demonstrated high accuracy in detecting and outlining lung tumors on CT scans, according to a study published in Radiology. This breakthrough could help doctors diagnose and treat lung cancer more efficiently, reducing the time and effort needed for manual tumor identification.
Lung cancer is the second most common cancer in men and women in the United States and remains the leading cause of cancer-related deaths, according to the American Cancer Society.
Detecting and tracking tumors on CT scans is crucial for monitoring cancer progression, evaluating treatment effectiveness, and planning radiation therapy. Currently, this process requires skilled radiologists to manually identify and outline lung tumors, a time-consuming task that varies from one physician to another.
AI-powered deep learning models have been introduced to assist in tumor detection, but previous models have struggled due to small datasets, reliance on manual adjustments, and limitations in segmenting multiple tumors.
To address these issues, researchers developed a more advanced deep learning model trained on one of the largest datasets of lung cancer CT scans ever used for this type of research.
The study, led by Dr. Mehr Kashyap from Stanford University School of Medicine, used 1,504 CT scans with 1,828 segmented lung tumors to train a deep learning model known as 3D U-Net.
The model was then tested on 150 CT scans, comparing its tumor identification with results from human physicians. The goal was to determine whether the AI could match or even improve upon the accuracy of expert radiologists.
The results were promising. The model correctly identified lung tumors in 92% of cases and had an 82% specificity rate, meaning it was good at avoiding false positives. When comparing how well the AI and physicians outlined tumors, the model’s accuracy was very close to human performance.
For 100 scans containing a single lung tumor each, the AI’s segmentation was 77% similar to physician-drawn outlines, while physician-to-physician comparisons reached 80% similarity. This suggests the model can segment tumors almost as well as human experts, while also saving time.
One of the key strengths of the model comes from using a 3D approach instead of a traditional 2D method. Many past AI models analyzed individual slices of a scan, which sometimes led to misidentifications, confusing tumors with blood vessels or airways.
In contrast, the 3D U-Net model processes full CT scans, allowing it to recognize smaller tumors and make more consistent decisions across different images.
Despite its success, the model does have some limitations. It sometimes underestimates tumor size, particularly when dealing with very large tumors, which could affect treatment planning.
Because of this, Dr. Kashyap recommends that the model be used in a physician-supervised workflow, allowing doctors to review and adjust AI-generated results as needed.
Moving forward, the researchers plan to explore additional applications of this AI tool. Future studies could test whether the model can track tumor changes over time, helping doctors monitor cancer progression and treatment response.
Another area of interest is whether the AI can predict patient outcomes based on tumor size and shape when combined with other medical data.
Dr. Kashyap believes this study is a significant step toward automating lung cancer diagnosis. AI models like this one could be integrated into treatment planning, tumor monitoring, and even personalized medicine, making lung cancer care more efficient and accessible in the future.
For more information about cancer, please see recent studies about the link between dairy food and certain cancers and this common food chemicals may cause cancer.
For more information about cancer, please see recent studies that plant-based diets may reduce risk of colorectal cancer in men, and Low-fat diet may help stop cancer growth.
The research findings can be found in Radiology.
Copyright © 2025 Knowridge Science Report. All rights reserved.
