Auto-segmentation And Quantitative Analysis Of Hepatic Lobules Based On Deep Learning

The hepatic lobule is the basic structure and functional unit of liver.The complex structure of hepatic lobule is composed of hepatic sinusoid,hepatic plate,peri-sinusoidal space and bile tubule with the central vein as the central axis.The anatomical structure of the liver has been known for hundreds of years,but the spatial relationship between the intricate vascular system and the hepatic lobule in the liver remains to be studied.The segmentation of hepatic lobule is of great significance for analyzing the morphological structure of hepatic lobules and studying its spatial distribution in the liver.Due to the difficulty of data acquisition of hepatic lobular images and the complexity of hepatic lobular labeling,there is a lack of a large number of accurate labeled data.The boundary of the hepatic lobule is blurred,the shape is irregular and the quality of the imaging data is not fine enough,which make it difficult to segment the hepatic lobule.For the new and complex data,the effect of traditional image segmentation method is not ideal.This study intends to use convolution neural network to segment hepatic lobules automatically in order to improve the efficiency of image analysis and reduce the workload of biological researchers or doctors.Aiming at the confocal imaging data of mouse liver,on the basis of accurate labeling of hepatic lobules and making full use of the prior knowledge of hepatic lobular structure,a set of hepatic lobular preprocessing method was established.And a segmentation model combined with multi-scale feature prediction fusion and attention mechanism was designed.The multi-scale module uses the global information on different scales to aggregate the features generated by skip connection multiple intermediate layers to predict.The attention mechanism eliminates the irrelevance in the skip connection and the ambiguity caused by the noise response,and increases the sensitivity of the model to the foreground pixels.The MIo U of the model on the hepatic lobular data set is 0.7161,and the average Dice coefficient is 0.8254.Compared with several mainstream segmentation networks,it achieves better segmentation performance with fewer parameters.The two-dimensional hepatic lobular data were reconstructed and rendered,and the hepatic lobules of automatic segmentation and manual segmentation were quantitatively analyzed respectively.The statistical results show that the average volume of 191 hepatic lobules automatically segmented is 0.0788±0.0042 mm~3,and the average surface area is 1.18±0.05 mm~2.The average volume of 150 hepatic lobules manually segmented is0.0674±0.0036 mm~3,and the average area is 1.09±0.04 mm~2.The shape of hepatic lobule is described by three quantitative parameters:sphericity,oblate ellipsoid and prolate ellipsoid.The sphericity is concentrated in 0.6-0.8,and the oblate ellipsoid is much larger than prolate ellipsoid.The shape of hepatic lobule tends to be irregular flat polygonal prism,similar to pebble.Compared with the quantitative results of manual segmentation,there was no significant difference in the volume(P=0.2230)and area(P=0.6037)of hepatic lobules calculated by two methods.The shape of the automatically segmented hepatic lobule is more irregular,which may be related to the imprecise segmentation of the edge of the hepatic lobule.Generally speaking,the automatic segmentation method not only improves the efficiency of hepatic lobule segmentation,but also ensures the accuracy of data analysis.