Optimization of an algorithm for IVIM parameter estimation from DWI-MRI in early diagnosis of hepatocellular carcinoma

Hepatocellular carcinoma (HCC) is the most common malignant liver tumor, characterized by reduced molecular diffusion, due to the high cellular density of the tumoral tissue, and marked hypervascularization, associated with increased arterial blood supply. The early detection of HCC is fundamental for effective and timely treatment, to prevent disease progression and improve the prognosis. For this purpose, Intravoxel Incoherent Motion Diffusion-Weighted Imaging (IVIM-DWI) represents a promising approach to quantify separately the pure molecular diffusion (D) and the microvascular perfusion (D* and f) components, without using contrast agents. Numerous studies have demonstrated the potential of IVIM-DWI in detecting and characterizing liver lesions. However, its clinical application remains challenging due to the lack of a standardized acquisition and processing protocol. This study aimed to optimize an algorithm for robust IVIM parameters quantification within liver, and to evaluate the effectiveness of an accelerated DWI acquisition protocol based on deep learning image reconstruction. Among the analysed fitting methods, the bi-exponential model implemented in the OSIPI repository, based on non-linear least squares estimation, provided the best compromise between accuracy and computational efficiency. The algorithm was then optimized and applied to DWI images. The quantitative analysis comparing IVIM parameters between HCC and liver parenchyma was performed on a retrospective dataset of 31 patients. The data were acquired using a 11 b-values DWI protocol on a 1.5 MR system (Siemens Healthineers, Forchheim, Germany). Summary metrics of IVIM parameters were extracted and compared between lesions and surrounding liver parenchyma, considering only small lesions (<3 cm, n=14). The reliability of parameter estimates and the quality of the fitting were assessed using the coefficient of variation (CV) of estimates and the residual sum of squares error (RSSE). A prospective data of 13 patients was used to evaluate the effectiveness of an accelerated DWI acquisition protocol using deep learning-based image reconstruction, compared to the standard protocol. Both protocols used 11 b-values. The study highlighted that DWI image quality is a key factor for robust and reliable IVIM quantification. Despite algorithm optimization, the fitting process remains highly sensitive to noise and motion artifacts. D* confirmed to be the least reliable parameter, showing high CV and large inter- and intra-patient variability, while D remained the most stable. Although no statistically significant differences were observed between HCC and liver parenchyma, a trend toward lower D and f values within lesions emerged, reflecting restricted diffusion and smaller overall perfusion fraction. Moreover, the reliability of estimates was significantly lower within HCC. In this context, deep learning-based image reconstruction proved to be especially promising, allowing to halve the acquisition time without compromising DWI image quality, and improving the reliability of IVIM estimation in liver. These preliminary results suggest the potential integration of deep learning-reconstruction into future IVIM protocols to achieve faster and more robust acquisitions. The findings of this study confirm the potential of IVIM-DWI as a contrast-free screening tool in liver imaging, but also highlight its strong dependance on image quality and acquisition protocol. Future studies should focus on validating the results on a larger and more heteogroneus cohort of patients, and on optimizing b-value distribution, minimizing motion artifacts during acquisition, and integrating deep learning-reconstruction for enhanced performance.