51 / 2025-03-14 10:15:28

Automated delineation of pyogenic liver abscess and prediction of Klebsiella pneumoniae infection based on CECT images with deep learning: a multicentre cohort study

Klebsiella pneumoniae,Pyogenic liver abscess,Radiomics,Deep learning,Prediction model

Background: Accurately identifying the bacterial strains causing Pyogenic liver abscess (PLA) is challenging in low-middle-income regions with limited resources. This study aimed to develop an early prediction model for Klebsiella pneumoniae liver abscess (KPLA) using automatic segmentation techniques.

Methods: Patients with PLA were enrolled from three medical centers between April 2011 and December 2023. The Pyogenic Liver Abscess Detection with AI (PLADA) model was developed based on the V-Net deep neural network. The contrast-enhanced computed tomography (CECT) images and clinical data of PLA patients were collected, and radiomics features were extracted and ranked based on their importance. Six machine-learning algorithms were employed to construct clinical, radiomics, and clinical-imaging models, which were further validated in an external cohort.

Results: 492 PLA patients were identified for analysis, 381 were in the training cohort and 111 in the external validation cohort. The PLADA model accurately segments the lesions with an average Dice coefficient of 0.941. The clinical-imaging model achieved a mean area under the curve (AUC) over 0.9 across all six algorithms in the training cohort, with Adaboost algorithm performed best in the external validation cohort (AUC of 0.847), and exhibited the highest sensitivity (79.8%) and specificity (81.8%) for diagnosing KPLA. The decision curve analysis (DCA) curve demonstrates its substantial clinical utility.

Conclusions: The PLADA model demonstrated high accuracy in segmenting PLA lesions. The clinical-imaging model, based on PLADA automatic segmentation, has the potential to reduce reliance on laboratory cultures and improve treatment accuracy for PLA patients.