The Study Of Segmentation Method In HCC Tumors Based On Multi-parametric MRI And Deep Learning

Hepatocellular Carcinoma(HCC)is one of the most common cancers all over the world.The accurate segmentation of lesions in medical images is very important for the determination of the tumor boundary and the volume measurement of the tumors.Magnetic resonance imaging(MRI)has become one of the main examinations of HCC because of its non-invasive nature and high resolution of soft tissue.Multi-parametric MRI images can provide tumor information in different ways,and the combination of these information may be more accurate in identifying the tumor lesions.However,manual delineation is a laborintensive procedure.Besides,the different experience of the clinicians may cause subjectivity in tumor contouring.In recent years,machine learning,especially deep learning,has made great breakthroughs in the field of medical image processing.Deep learning algorithm can automatically learn the tumor features and mine potential high-dimensional semantic information,so as to achieve automatic recognition and segmentation of tumor lesions.Thus we aimed to apply deep learning in HCC tumor segmentation by using multiparametric MRI.We built a multi-input model of deep network to achieve accurate,quick and objective HCC tumor segmentation.To study the influence of different inputs on the segmentation results,we tested the segmentation results based on single sequence MRI and multi-parametric MRI respectively.First,according to the suggestion of the clinicians,we proposed a deep convolutional neural network(DCNN)scheme based on the Gd-EOB-DTPA enhanced MRI(EOB)only.To improve the low recognition rate of small tumors,we improved the deep network architecture,which preserved the spatial resolution of EOB images and the scale information of small and medium-sized tumors.We collected EOB images from 51 patients and completed leave-oneout cross-validation.The average Dice coefficient(DSC)of the segmentation tasks for these 51 HCC cases was 0.722,which was better than the existed studies.The average sensitivity was 0.708 and the average precision was 0.792.Second,we fused the information of EOB images and portal phase MRI images based on deep fused network(DFN)for the segmentation of HCC tumor lesions.The purpose was to reduce the false positive and false negative segmentation through the complementation of different MRI images.To prevent DFN from overfitting,we designed a random data training strategy to train each sub module of the network.In order to explore the mechanism of DFN in processing two kinds of MRI information,we designed a single input experiment for each MRI sequence respectively.It was proved that the input of multi-parametric MRI images could suppress false positive and avoid false negative.Similarly,we completed leave-oneout cross-validation with the dataset of 51 patients.The average DSC of the segmentation tasks for these 51 HCC cases achieved 0.771,better than that in single-parametric MRI segmentation results.The average sensitivity was 0.751 and the average precision was 0.803.The above results showed that deep learning method performed on multi-parametric MRI images achieved better segmentation than that based on single-parametric MRI images.However,our method was implemented on two-dimensional images and was unable to make use of the spatial information of three-dimensional data.In addition,only the images with tumor lesions were used for training and testing.The segmentation performance on MRI images without tumor lesions has not been verified.In the future studies,a 3-dimensional DCNN method can be considered for HCC tumor segmentation.All MRI images(with or without tumor lesions)may be used in order to increase the robustness of the algorithm.If our proposed method is validated in future studies,it will be of imp ortance in assisting the clinical diagnosis and treatment of HCC.