# Assessment of deep learning reconstruction effects on detection and differentiation of liver metastasis from hepatic hemangioma in diffusion-weighted imaging

## TL;DR

This study compares deep learning-enhanced and traditional imaging methods for detecting liver tumors and hemangiomas, finding improved image quality but similar diagnostic accuracy.

## Contribution

The study evaluates deep learning reconstruction's impact on liver metastasis and hemangioma differentiation in diffusion-weighted imaging.

## Key findings

- DL-DWI showed better metastasis conspicuity and higher SNR/CNR compared to CS-DWI.
- ADC values of liver and lesions were lower in DL-DWI than in CS-DWI.
- Diagnostic performance metrics like ROC area, sensitivity, and specificity were comparable between DL-DWI and CS-DWI.

## Abstract

To evaluate and compare the performance of diffusion-weighted imaging (DWI) using compressed sensing (CS) and DWI using CS with model-based deep learning reconstruction (DL-DWI) in detecting and differentiating liver metastases from hepatic hemangiomas.

We retrospectively analyzed data from 53 patients with metastases or hemangiomas (34 men and 19 women, mean age, 65.9 years) who underwent abdominal DWI. Two radiologists evaluated liver contour and distortion, artifact, noise, overall image quality, and lesion conspicuity using a five-point scale. Signal-to-noise ratio (SNR) and apparent diffusion coefficient (ADC) of the liver, as well as contras-to-noise ratio (CNR) and ADC of metastases (n = 59) and hemangiomas (n = 33) were assessed and statistically compared. A receiver operating characteristic (ROC) analysis was performed to assess the diagnostic performance of the two sequences for differentiating metastases and hemangiomas.

DL-DWI provided significantly better conspicuity of metastasis than CS-DWI (p < 0.05 in both radiologists), whereas no significant difference was observed in the conspicuity of hemangioma between DL-DWI and CS-DWI. The SNR of liver parenchyma and the CNR of metastases and hemangiomas were higher in DL-DWI than in CS-DWI (p < 0.05). ADC values of liver parenchyma, metastases, and hemangiomas were lower in DL-DWI than in CS-DWI (p < 0.05). The ADC cutoff value for differentiating between metastases and hemangiomas was 1.693 × 10–3 mm2/s in DL-DWI and 1.411 × 10–3 mm2/s in CS-DWI. No significant differences were observed in the area under the ROC curve, sensitivity, and specificity between the two methods (p > 0.05).

DL-DWI enhanced both qualitative and quantitative aspects of image quality in abdominal DWI. However, its diagnostic performance, including ADC cutoff values for differentiating between metastases and hemangiomas, is comparable to that of CS-DWI.