In a cohort study reported in JAMA Network Open, Azari et al evaluated whether machine learning–guided analysis of intraoperative molecular imaging (IMI) data could accurately and rapidly determine the malignant potential of indeterminate lung nodules during surgery. The study was undertaken to address persistent challenges in localizing small pulmonary nodules, and the time, cost, and error-prone nature of intraoperative frozen section analysis. By integrating artificial intelligence (AI)-based image segmentation with statistically derived nomograms, the investigators sought to develop an “optical biopsy” approach capable of assisting real-time surgical decision-making.
Study Details
Data were analyzed retrospectively from a prospectively collected database of patients with indeterminate lung nodules treated at the University of Pennsylvania between 2014 and 2021. A total of 322 patients were included in the study, with 279 having complete clinical and imaging data suitable for algorithm development and validation. The cohort was predominantly female (62.7%), and all patients were considered to have high-risk nodules potentially harboring malignancy.
All patients underwent IMI-guided lung surgery to highlight malignant tissue intraoperatively. Fluorescence intensity from tumor tissue and surrounding normal lung parenchyma was quantified to calculate the tumor-to-background ratio (TBR), a key imaging metric. Patients were randomly divided into training and validation sets in an 8:2 ratio, with extensive cross-validation to reduce selection bias.
The research team developed a machine learning–based image segmentation algorithm capable of rapidly and reproducibly calculating TBR directly from intraoperative images. This image analyzer was integrated with nomograms to create a unified optical biopsy algorithm. The combined system was then tested retrospectively in the validation cohort and prospectively in an independent cohort of 61 consecutive patients undergoing IMI-guided lung cancer surgery.
Key Findings
In retrospective analyses, two related nomogram models demonstrated strong discriminatory performance for malignant vs benign nodules, with areas under the receiver operating characteristic curve ranging from 0.865 to 0.893. Variables significantly associated with malignancy included smoking history greater than 5 pack-years, elevated ex vivo TBR values, elevated TBR after bisection of the specimen, and detectable in situ fluorescence. The machine learning–based image analyzer produced TBR measurements comparable to manual calculations but with significantly less variability and markedly shorter processing times.
When integrated into the final optical biopsy algorithm, the system accurately estimated malignancy risk in the validation cohort, correctly classifying all benign lesions and 96% of invasive adenocarcinomas. Misclassification occurred in a small number of cases due to known technical limitations, such as fluorescence dampening from blood products.
In the prospective cohort, the algorithm demonstrated a sensitivity of 93.8%, specificity of 100%, positive predictive value of 100%, and negative predictive value of 71%. Notably, the algorithm generated results in less than 2 minutes on average, compared with a mean of 34 minutes required for frozen section analysis. False-negative results occurred primarily in patients with heavy smoking histories and marked anthracosis, which increased background fluorescence and reduced contrast between tumor and normal tissue.
The authors concluded: “In this cohort study of patients with indeterminate lung nodules, intraoperative imaging data analyzed by AI accurately determined if a nodule was malignant. This has the potential to improve the diagnostic challenges that occur at the time of surgery.”
Feredun Azari, MD, of Heart, Vascular, and Thoracic Institute, Cleveland Clinic Foundation, Cleveland, Ohio, is the corresponding author of the JAMA Network Open article.
DISCLOSURE: For full disclosures of the study authors, visit jamanetwork.com.
