A unified imaging-to-pharmacology AI framework for pulmonary nodule detection, malignancy classification, and mutation-stratified drug validation from low-dose CT scans

Abstract

Background: Low-dose computed tomography (CT) screening for lung cancer can impose a substantial and clinically under-recognised psychosocial burden, with published studies reporting clinically significant anxiety among 30%–45% of screen-positive individuals, driven by a 96.4% benign false-positive rate and prolonged diagnostic uncertainty. Chronic distress may activate the hypothalamic–pituitary–adrenal axis and contribute to catecholamine-driven β2-adrenergic–epidermal growth factor receptor transactivation. Methods: We present PulmoNet-X, a unified deep learning framework that integrates a six-module imaging backbone with three psychosomatic burden-reduction mechanisms: false-positive reduction module (FPRM), uncertainty-calibrated risk stratification (UCRS), and integrated drug validation and therapeutic drug recommendation (IDVTRM). The framework was evaluated on the Lung Image Database Consortium and Image Database Resource Initiative (LIDC-IDRI) benchmark comprising 888 CT scans and 1,004 annotated nodules using 10-fold cross-validation. Results: PulmoNet-X achieved 98.1% nodule-detection sensitivity at one false-positive/scan (competition performance metric = 93.7%), malignancy classification area under the curve (AUC) = 97.8%, segmentation Dice similarity coefficient (DSC) = 84.6%, and Lung Imaging Reporting and Data System agreement κ = 0.87 on LIDC-IDRI. FPRM reduced false-positive detections from 4.8 to 1.0 per scan, corresponding to a modelled estimate of approximately 380 fewer unnecessary recall investigations per 1,000 patients screened. UCRS achieved a calibration error of 0.016 (expected calibration error). IDVTRM attained a mutation-stratified drug-response AUC of 0.918 and lung-tissue penetration prediction (R2 = 0.986), with 94.2% concordance with independent expert oncologist assessment. All primary pairwise comparisons were statistically significant (two-sided Wilcoxon signed-rank test, p < 0.001, Bonferroni-corrected for six comparisons). Conclusion: External validation on 150 independent National Institutes of Health Chest CT scans provided preliminary evidence supporting generalisability beyond LIDC-IDRI (sensitivity = 98.1%, AUC = 99.3%, DSC = 95.8%, κ = 0.94). Reported psychosomatic benefits remain mechanistically grounded hypotheses; validated psychometric instruments, including the Generalised Anxiety Disorder 7-item scale, Patient Health Questionnaire-9, and Hospital Anxiety and Depression Scale, were not administered. The pharmacological prediction component remains exploratory, and the incremental contribution of imaging features to drug-response prediction has not been established. Prospective, multi-centre validation is required before clinical translation. PulmoNet-X demonstrates the feasibility of incorporating psychosomatic burden reduction as an explicit architectural objective alongside imaging and pharmacological modelling. Relevance for patients: PulmoNet-X is designed to reduce unnecessary recalls, screening anxiety, and diagnostic uncertainty by providing calibrated cancer risk estimates and personalised treatment guidance.