Machine Learning-Guided Detection of Malignancy of Lung Nodules With Molecular Imaging-Guided Surgery

Over 1 million pulmonary nodules are discovered each year in the US, and many of these undergo molecular imaging-guided surgery to obtain a diagnosis. Locating a small nodule and determining its malignant potential is technically challenging and is prone to human error.

To demonstrate use of a machine learning (ML) algorithm with molecular imaging to analyze imaging data during lung cancer surgery to determine malignant potential of nodules.

Data were retrospectively analyzed from a prospectively collected database. Between 2014 and 2021, patients at the hospital of the University of Pennsylvania with lung nodules were included in the study. Patients in the model development set were randomly allocated into training and validation sets in an 8:2 ratio. Data were analyzed from January 2014 and December 2021.

Algorithmic tumor to background ratio (TBR) detection was implemented for individual images using Image Processing Toolkit. Developed nomogram and artificial intelligence (AI) image analyzer were combined as an optical biopsy algorithm and tested prospectively between 2021 and 2024.

A total of 322 patients with lung nodules were included in the study, of whom 279 had complete clinical data for data analysis (175 [62.7%] female). The nomograms and image segmentation technology were developed using a large database of IMI videos (1014 video sequences) and demonstrated an area under the curve of 0.865 to 0.893 for malignant nodule assessment. On multivariate logistic regression analysis, patient smoking history of greater than 5 pack-years (patient pack-years [PPY] >5), ex vivo back table TBR greater than 2.0, ex vivo bisected tumor lesions TBR greater than 2.4, and in situ (inside the chest) fluorescence were found to have statistically significant associations with malignancy on final pathology. Prospective testing in an independent set of 61 consecutive patients during IMI-guided cancer surgery demonstrated a sensitivity of 93.8%, specificity of 100%, positive predictive value of 100%, and negative predictive value of 71%. The study algorithm determined malignant potential of the lesion in less than 2 minutes (mean [SD], 1.8 [0.17] minutes) compared with a mean (SD) of 34 (11) minutes with frozen section analysis.

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.