Codebook-guided multi-cue normality modeling for pulmonary CT anomaly detection.

INTRODUCTION

Automated anomaly detection from chest computed tomography (CT) may support research on pulmonary image triage without requiring exhaustive abnormal annotations. Supervised methods depend on costly expert labels, while many existing normal-only methods rely predominantly on a single anomaly cue. In pulmonary CT, heterogeneous lesion appearance and variation across slices and acquisition sources may reduce the reliability of single-cue scoring. This study develops a codebook-guided framework that integrates complementary cues for abnormality detection and spatial heatmap generation.

METHODS

The framework combines residual vector-quantized reconstruction, transformer-based token discrimination learned through codebook-guided latent perturbation, and quantization-distance scoring. The primary benchmark combined three public CT datasets. Slice-level evaluation used a visually screened, lesion-enriched subset, whereas patient-level evaluation retained all valid lung-containing slices and aggregated the highest 10% of slice scores by their mean. Patient-level thresholds were calibrated using normal validation patients only. Quantitative localization was additionally evaluated on 100 LIDC-IDRI cases containing nodules marked by at least three radiologists.

RESULTS

The proposed method achieved a slice-level AUROC of 0.9218. At the patient level, it achieved an AUROC of 0.9187 ± 0.0123 , balanced accuracy of 0.7521 ± 0.0354 , and F1-score of 0.9297 ± 0.0052 ; the patient-stratified bootstrap 95% confidence interval for AUROC was 0.8786-0.9727. Disease-wise patient-level AUROC ranged from 0.8470 for COVID-19 to 0.9747 for community-acquired pneumonia. On the external LIDC-IDRI localization cohort, the proposed method achieved a P-AUROC of 0.9155, the highest observed value among the evaluated methods.

DISCUSSION: The findings indicate that multi-cue normality modeling remains effective after patient-level aggregation and provides quantitative localization capability on an external nodule cohort. Same-source comparisons within COVID-CT-MD reduce, but do not eliminate, concern about dataset-source confounding. Because the study used retrospective public datasets and lacked prospective patient-level external validation, the results should be interpreted as methodological evidence rather than clinical validation.