Automated detection of pyogenic liver abscess and diagnosis of Klebsiella pneumoniae infection based on CECT images with deep learning: A multicenter study

Background

Accurate identification of bacterial pathogens in pyogenic liver abscess (PLA) remains challenging in low- and middle-income countries (LMICs). We developed and validated an artificial intelligence (AI) model that replicates clinical workflow to detect and diagnose Klebsiella pneumoniae liver abscess (KPLA) using contrast-enhanced computerized tomography (CECT).

Methods

Patients with PLA were enrolled from three medical centers between April 2011 and April 2024. The Pyogenic Liver Abscess Detection with AI (PLADA) system employed a V-Net deep neural network for automated lesion segmentation. Radiomics features were subsequently extracted and ranked by importance. Six machine learning algorithms were used to develop clinical, radiomics, and integrated clinical-imaging models for KPLA diagnosis in the training cohort, with external validation in a multicenter framework.

Results

A total of 492 PLA patients were identified. The PLADA algorithm accurately segments the lesions with a mean dice coefficient of 0.941 (95 % confidence interval: 0.899–0.983). The clinical-imaging model demonstrated strong discriminatory performance, with a mean area under the curve (AUC) of over 0.9 across all algorithms in the training cohort. The AdaBoost-based clinical-imaging model achieved an AUC of 0.847 (95 % CI: 0.753–0.940) in the external validation cohort, with high sensitivity (71.9 %) and specificity (86.4 %). Decision curve analysis (DCA) confirmed clinical utility.

Conclusions

The PLADA system provides an accurate, non-invasive method for KPLA diagnosis through a clinically interpretable two-stage AI framework. While limited by a retrospective, single-country dataset, this approach shows particular promise for early detection of KPLA in resource-limited settings.