Clinical application of deep neural networks in laser endomicroscopic molecular imaging for accurate assessment of lung nodules
Year: 2026
Type: article
Abstract: Needle-based near-infrared confocal laser endomicroscopy (NIR-nCLE) enables intraoperative imaging of pafolacianine-labeled cells within pulmonary lesions, but interpretation requires operator experience. We evaluated a machine-learning pipeline for classifying NIR-nCLE recordings as benign or malignant using final histopathology as the reference standard. Fifty patients undergoing pafolacianine-guided resection were prospectively enrolled, and 248 lesions were sampled. An AlexNet-based convolutional neural network classified individual frames, with lesions classified as malignant when the proportion of frames predicted to be malignant was ≥ 0.02. Lesion-level performance was assessed by stratified five-fold cross-validation using 121 analyzable recordings available in the retained de-identified archive, comprising 70 malignant and 51 benign lesions and 23,589 frames. All frames from each lesion were confined to a single fold; patient-level partitioning was not possible because patient-to-recording links were unavailable. Sensitivity was 98.6% (95% CI 95–100%), specificity was 72.5% (60–84%), positive predictive value was 83.1%, negative predictive value was 97.4%, accuracy was 87.6% (81–93%), and the area under the receiver operating characteristic curve (AUROC) was 0.97 (0.94–0.99). Logistic regression using mean fluorescence intensity alone achieved an AUROC of 0.97 (0.94–0.99) and an accuracy of 90.1%; a 100-unit feature-based neural network achieved an AUROC of 0.98 (0.95–1.00). In an internal exploratory intraoperative cohort of 10 additional patients, all nine malignant and 18 benign lesions were correctly classified before final pathological evaluation, yielding a sensitivity of 100% (95% CI 66.4–100%) and a specificity of 100% (81.5–100%). Automated classification of NIR-nCLE recordings demonstrated promising preliminary internal performance, with similar AUROC estimates for simpler fluorescence-based models. Independent external validation with patient-level separation is required before clinical use.
Source: Scientific Reports
Language: English
Topic: AI in cancer detection
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,690
OpenAlex ID: W7225645737
FWCI:
Cites:
Cited by:
Related to: