Development and internal validation of a multimodal deep learning model integrating imaging, clinical and pathological features to predict false-negatives in non-specific benign lung biopsies: a dual-center study.

Authors

Abstract

Computed tomography (CT)-guided percutaneous transthoracic needle biopsy (PTNB) is widely used for lung lesion evaluation. Non-specific benign PTNB results create important clinical challenges. These outcomes carry a considerable risk of false-negative malignancy, with reported rates as high as 49%. Missed cancer diagnoses may delay curative-intent treatment. This study developed and internally validated a multimodal deep learning model. The model combines CT imaging, clinical variables, and pathological features. Its goal was to distinguish false-negative malignancy from non-specific benign PTNB results. In this dual-center retrospective study, consecutive patients with non-specific benign PTNB results from two tertiary hospitals (January 2020-December 2022) were screened, and 403 patients (mean age, 57.19±10.2 years; 289 male patients) with complete clinical, imaging, and follow-up data were included. The reference standard for the final diagnosis combined at least 2 years of imaging stability or resolution, surgical pathology, repeat biopsy, or documented clinical progression. Candidate predictors included CT imaging features, clinical variables (demographics, smoking status, tumor markers), and four pathological descriptors. Three models were developed: an imaging model (CT only), an imaging-clinical model (CT and clinical data), and a combined model (CT, clinical data, and pathological features). Patients were randomly split into training (80%) and test (20%) sets. Model performance was evaluated using receiver operating characteristic (ROC) analysis and compared using the DeLong test. Diagnostic accuracy was also compared with that of the four thoracic radiologists (4-15 years of experience). Overall, 90 patients (22.3%) had false-negative malignancy, with 19 events in the test set (n=78). The combined model achieved an area under the curve (AUC) of 0.825 [95% confidence interval (CI): 0.710, 0.923], significantly outperforming the imaging model (AUC 0.655, 95% CI: 0.507, 0.794; P=0.04). It was numerically superior to the imaging-clinical model (AUC, 0.789; 95% CI: 0.667, 0.892; P=0.26), with no statistically significant difference observed between the two models. At the optimal threshold of 0.225, the combined model showed an accuracy of 0.782 (95% CI: 0.678, 0.859), a sensitivity of 0.842 (95% CI: 0.624, 0.945), and a specificity of 0.763 (95% CI: 0.640, 0.853), noting that 75.6% of test cases were true negatives. The combined model's accuracy exceeded that of the four radiologists (range, 58.97-73.08%), although this comparison was descriptive without formal statistical testing. This multimodal model shows promising discriminative performance for predicting false-negative PTNB results. Since performance metrics and the 0.225 threshold were obtained from internal random hold out validation, external independent-cohort validation and threshold optimization are mandatory prior to clinical decision-support application.

Tags

Topics