A novel deep learning model shows potential for detecting and segmenting lung cancers, according to a study published in Radiology. The study’s findings could have significant consequences for treating lung cancer.
Model and clinical segmentation examples. (A) 71-year-old female with non-small cell lung cancer (NSCLC) from the internal test set. (B) 87-year-old male with NSCLC from the external test set. Both patients underwent radiotherapy. CT simulation scans acquired prior to radiotherapy are displayed. Clinical segmentations represent the contours developed in the radiotherapy planning process. Image Credit: RSNA
According to the American Cancer Society, lung cancer is the second most frequent disease in both men and women in the United States, and it is the leading cause of cancer death.
Accurate detection and segmentation of lung tumors on CT images is crucial for tracking cancer growth, assessing treatment efficacy, and planning radiation therapy. Currently, qualified physicians manually identify and segment lung cancers on medical images, a time-consuming process susceptible to physician variability.
Although AI deep learning techniques have been used for lung tumor detection and segmentation, previous research has been constrained by small datasets, manual input, and a focus on lung tumor segmentation alone. This underscores the need for models that can reliably and automatically delineate tumors in various clinical contexts.
This study developed a near-expert-level lung tumor identification and segmentation model using a unique, large-scale dataset of routinely collected pre-radiation treatment CT simulation scans and their corresponding clinical 3D segmentations. The main goal was to create a model that reliably recognizes and separates lung cancers on CT scans from various medical facilities.
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.
Mehr Kashyap, MD, Study Lead Author and Resident Physician, Department of Medicine, Stanford University School of Medicine
Model Architecture Provides Advantage
The retrospective study used 1,504 CT scans containing 1,828 segmented lung tumors to train an ensemble 3D U-Net deep-learning model for lung tumor detection and segmentation. After that, 150 CT scans were used to test the model.
Tumor volumes predicted by the model and those defined by the physician were compared. Sensitivity, specificity, false positive rate, and dice similarity coefficient (DSC) were among the performance indicators. The model segmentations were compared with those from the three physician segmentations to get the model-physician DSC values for each pairing.
Using the combined 150-CT scan test set, the model detected 82% specificity and 92% sensitivity of lung tumors.
In a group of 100 CT scans with a single lung tumor, the median model-physician and physician-physician segmentation DSCs were 0.77 and 0.80, respectively. The model required less segmentation time than physicians.
Dr. Kashyap feels that employing a 3D U-Net architecture to construct the model gives it an advantage over alternatives using a 2D architecture.
He added, “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.”
The model’s propensity to underestimate tumor volume was one of its drawbacks; this led to worse performance on really large tumors. As a result, Dr. Kashyap advises that the model be used in a workflow that is overseen by a physician. This will enable physicians to detect and eliminate lower-quality segmentations and erroneously identified lesions.
According to the researchers, future studies should concentrate on using the model to calculate the overall burden of lung tumors and assess treatment response over time, comparing it with current techniques. Additionally, they advise evaluating the model’s capacity to predict clinical outcomes using estimated tumor burden, especially when paired with other prognostic models that use a variety of clinical variables.
“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,” Dr. Kashyap concluded.
Journal Reference:
Kashyap, M., et. al. (2025) Automated Deep Learning-Based Detection and Segmentation of Lung Tumors at CT. Radiology. doi.org/10.1148/radiol.233029