Deep learning-based image reconstruction algorithm for lung diffusion weighted imaging: improved image quality and diagnostic performance

To assess the impact of deep learning reconstruction (DLR) on the image quality and the diagnostic performance of lung DWI.

Totally 46 patients with 46 lesions (malignant 35, benign 11) were prospectively recruited and imaged with DWI. DWI images were reconstructed with conventional reconstruction (ConR) and DLR, respectively. Two radiologists evaluated the signal-to-noise ratio (SNR), apparent diffusion coefficient (ADC) and IVIM-derived parameters of pulmonary lesions. Intraclass correlation coefficients were used to evaluate reader agreement. Paired t-tests or Wilcoxon tests were used to compare the measurements between DLR and ConR. Differences between different types of lesions were determined using Student’s or Mann-Whitney test. ROC curves were used to evaluate the diagnostic performance of the DWI parameters.

All the measurements with DLR and ConR showed excellent inter-observer consistency. In comparison with ConR, DLR showed higher SNR, higher measurements of ADC and D_slow (all P < 0.001). ADC and f_fast of malignant lesions was significantly lower than that of benign lesions with DLR and ConR (P < 0.05), respectively, but no difference was found between the two reconstruction algorithms. ROC analysis showed the diagnostic performance of ADC and f_fast with DLR outweighed ConR slightly for distinguishing malignant from benign lesions. The ADC of DLR can distinguish adenocarcinoma from squamous cell carcinoma significantly (P < 0.001) with AUC 0.734.

DLR can significantly increase the SNR of lung DWI which affects the DWI quantification, resulting with improved diagnostic performance for distinguishing benign from malignant lesions, especially discriminating adenocarcinoma from squamous cell carcinoma.