Deep learning brings automated readout to lung nodule molecular imaging
Lung nodules remain one of thoracic oncology’s most demanding diagnostic problems because visual appearance alone often fails to separate benign tissue from malignancy with confidence. Intraoperative needle-based near-infrared confocal laser endomicroscopy (NIR-nCLE) paired with pafolacianine offers a way to inspect labeled cells in real time, but the signal is only useful if someone can interpret it quickly and accurately while the patient is still on the table. This study asks whether a machine-learning pipeline can shoulder part of that burden.
The answer, within the limits of the internal data set, is encouraging. A convolutional neural network based on AlexNet classified individual frames from NIR-nCLE recordings, and lesion-level calls were made when at least 2% of frames were predicted malignant. On cross-validated analysis of 121 analyzable recordings, the model reached 98.6% sensitivity and 72.5% specificity, with an AUROC of 0.97. In a small internal intraoperative cohort, all nine malignant lesions and all 18 benign lesions were classified correctly before final pathology was available.
Frame-level signal, lesion-level decision
The study enrolled 50 patients undergoing pafolacianine-guided resection and sampled 248 lesions. Only 121 recordings were retained in the de-identified archive used for the main lesion-level analysis, comprising 70 malignant and 51 benign lesions and 23,589 frames. Every frame from a given lesion was kept within a single fold during stratified five-fold cross-validation, but patient-level partitioning could not be performed because patient-to-recording links were unavailable. That design detail matters because frame-level leakage can inflate apparent performance when adjacent images from the same lesion are split across training and testing sets.
Even with that constraint, the signal was strong. Sensitivity of 98.6% suggests the model rarely missed malignant lesions in this curated internal set. Specificity of 72.5% shows a different part of the problem: benign lesions were more likely to be mislabeled than malignant ones were to be overlooked. In practical terms, that balance favors a rule-out tool more than a definitive rule-in tool, at least in its current form.
One sentence captures the clinical tension: the model was far better at avoiding false negatives than at eliminating false positives.
Fluorescence intensity may carry most of the predictive weight
The most striking result is not the neural network alone. Logistic regression using mean fluorescence intensity achieved an AUROC of 0.97 and an accuracy of 90.1%, while a 100-unit feature-based neural network reached an AUROC of 0.98. Those figures place the deep learning approach in the context of a simpler fluorescence-based measurement that performed nearly as well on AUROC.
That comparison is important because pafolacianine imaging is fundamentally a molecular contrast problem. If fluorescence intensity already captures much of the malignant-versus-benign separation, then the added value of a more complex model may lie in pattern recognition at the frame and lesion level rather than in discovering a hidden biological signature. In the operating room, a simpler and more transparent model may be easier to trust, audit, and deploy, especially if it performs comparably to a convolutional neural network.
The authors’ intraoperative exploratory cohort reinforces the possibility of real-time use. Among 10 additional patients, nine malignant and 18 benign lesions were correctly classified before final pathological evaluation, yielding 100% sensitivity and 100% specificity in that small sample. The confidence intervals remain wide, and the sample is too small to support clinical adoption, but the observation suggests the workflow can operate in a live surgical environment rather than only on retrospective data.
Why the biology fits the machine-learning approach
NIR-nCLE images are not ordinary surgical photographs. They are molecular readouts of labeled tissue, which means the imaging surface already contains biologically enriched contrast. Malignant lesions may differ from benign lesions in cellular uptake, fluorescence distribution, and structural organization, all of which can produce frames that are visually distinct even before a human viewer assigns meaning to them. Deep neural networks are well suited to exactly this kind of task: they can convert subtle, repeatable differences in pixel-level structure into a classification decision without requiring a clinician to identify every feature manually.
In this setting, the model’s high sensitivity is clinically coherent. When a device can identify most malignant lesions from a fluorescence-based signal, it has the potential to reduce the chance that an occult cancerous focus is left behind because the imaging impression was too ambiguous in the moment. The lower specificity is the familiar price of a conservative detector: some benign lesions will trigger malignant predictions, and a surgeon or pathologist still needs a confirmatory framework before committing to a management decision.
The main limitation is not mathematical, but evidentiary
The strongest caution in this paper is not the algorithm itself. It is the validation structure. The main analysis used a retained de-identified archive, and patient-level separation was not possible because links between patients and recordings were unavailable. That prevents a fully clean estimate of generalizable performance. Independent external validation with patient-level partitioning is still required, and the authors say so plainly.
The study also leaves open whether the deep model truly exceeds simpler fluorescence-derived methods in a clinically meaningful way. If mean fluorescence intensity alone can deliver comparable AUROC, then the bar for adoption rises: a more complex neural network will need to prove not just accuracy, but robustness, portability, and operational value. Those are different claims.
Even so, the message from this early internal work is hard to ignore. Automated interpretation of NIR-nCLE recordings appears feasible, and the bedside use case is obvious: rapid, intraoperative assessment of lung nodules when decisions are being made under time pressure and tissue is scarce. Whether that promise survives prospective external testing will determine if this approach becomes a research tool, a decision-support adjunct, or something more durable in thoracic oncology practice.