Deep Learning Model Detects Lung Tumors on CT
Posted on 27 Jan 2025
Lung cancer is the second most common cancer in the U.S. and the leading cause of cancer-related death. Accurately detecting and segmenting lung tumors on CT scans is essential for monitoring progression, evaluating treatment, and planning radiation therapy. Currently, this process is manually performed by clinicians, which is labor-intensive and prone to physician variability. While AI deep learning methods have been applied to tumor detection and segmentation, previous studies have been limited by small datasets, manual inputs, and a focus on single tumors. These limitations highlight the need for models that can provide robust, automated tumor delineation in various clinical settings. Now, a new study published in Radiology has shown promise for a new deep learning model capable of accurately detecting and segmenting lung tumors, positively impacting lung cancer treatment.
In this study, researchers at Stanford University School of Medicine (Stanford, CA, USA) used a large-scale dataset from routine pre-radiation treatment CT scans to develop a near-expert-level model for lung tumor detection and segmentation. Their aim was to create a model capable of identifying and segmenting lung tumors across different medical centers. The team used an ensemble 3D U-Net deep learning model trained on 1,504 CT scans with 1,828 segmented tumors and tested it on 150 CT scans. The model's predictions were compared with physician-delineated tumor volumes. Key performance metrics, such as sensitivity, specificity, false positive rate, and Dice similarity coefficient (DSC), were calculated to assess model accuracy. The model achieved 92% sensitivity and 82% specificity in detecting lung tumors. For scans with a single tumor, the median model-physician DSC was 0.77, and the physician-physician DSC was 0.80. The model also performed faster than physicians.
The researchers believe that the 3D U-Net architecture provides advantages over 2D models by capturing interslice information, which helps identify smaller lesions that may be confused with other structures. However, the model tended to underestimate tumor volume, especially for larger tumors, which could affect performance. The researchers recommend integrating this model into a physician-supervised workflow to allow clinicians to validate and correct any misidentified lesions. They also suggest future research to evaluate how the model can be applied to assess overall lung tumor burden and treatment responses over time, and whether it can predict clinical outcomes in combination with other prognostic models.
"To the best of our knowledge, our training dataset is the largest collection of CT scans and clinical tumor segmentations reported in the literature for constructing a lung tumor detection and segmentation model," said the study's lead author, Mehr Kashyap, M.D. "Our study represents an important step toward automating lung tumor identification and segmentation. This approach could have wide-ranging implications, including its incorporation in automated treatment planning, tumor burden quantification, treatment response assessment and other radiomic applications."
Related Links:
Stanford University School of Medicine
