Liver and Focal Lesion Segmentation in Multi-Sequence MRI Enhanced by Cascaded Deep Learning

Automatic segmentation of medical images is considered to be one of the most promising applications of deep learning models. However accurate segmentation of liver lesions in MRI remains an open problem. This study presents a cascaded deep learning approach for liver and focal lesion segmentation on multi-sequence MRI, specifically using Diffusion-Weighted Imaging (DWI) and T2-weighted sequences. The baseline method, utilizing a U-Net architecture, achieved robust liver segmentation but faced challenges in accurately segmenting lesions, with mean Dice Similarity Coefficients (DSC) of 0.700 and 0.656 on DWI and T2, respectively. To address this, a novel cascaded prediction technique was introduced, wherein segmentation predictions from one modality were iteratively refined by incorporating information from the other modality. This approach significantly improved lesion segmentation performance, with mean DSC increasing to 0.754 on DWI and 0.714 on T2 after three cascade steps. The model also showed substantial gains in lesion-wise sensitivity, crucial for accurate detection, increasing from 0.734 to 0.790 for DWI and from 0.717 to 0.753 for T2. Validation on the LLD-MMRI dataset further demonstrated the method’s generalization potential, achieving a lesion-wise sensitivity of 0.723 for DWI and 0.761 for T2. These findings highlight the effectiveness of the cascaded model in improving segmentation accuracy and suggest its utility in clinical applications for liver and lesion analysis in MRI.