Convolutional neural networks for semantic segmentation of FIB-SEM volumetric image data

Detta är en Master-uppsats från Göteborgs universitet/Institutionen för matematiska vetenskaper

Sammanfattning: Focused ion beam scanning electron microscopy (FIB-SEM) is a well-established microscopytechnique for 3D imaging of porous materials. We investigate three poroussamples of ethyl cellulose microporous films made from ethyl cellulose and hydroxypropylcellulose (EC/HPC) polymer blends. These types of polymer blends are usedas coating materials on various pharmaceutical tablets or pellets and form a continuousnetwork of pores in the film. Understanding the microstructures of these porousnetworks allow for controlling drug release. We perform semantic segmentation of theimage data, separating the solid parts of the material from the pores to accuratelyquantify the microstructures in terms of porosity. Segmentation of FIB-SEM data iscomplicated because in each 2D slice there is 2.5D information, due to parts of deeperunderlying cross-sections shining through in porous areas. The supposed shine-througheffect greatly complicates the segmentation in regards to two factors; uncertainty inthe positioning of the microstructural features and overlapping grayscale intensitiesbetween pore and solid regions.In this work, we explore different convolutional neural networks (CNNs) for pixelwiseclassification of FIB-SEM data, where the class of each pixel is predicted using athree-dimensional neighborhood of size (nx; ny; nz). In total, we investigate six typesof CNN architectures with different hyperparameters, dimensionalities, and inputs.For assessing the classification performance we consider the mean intersection overunion (mIoU), also called Jaccard index. All the investigated CNNs are well suitedto the problem and perform good segmentations of the FIB-SEM data. The so-calledstandard 2DCNN performs the best overall followed by different varieties of 2D and3D CNN architectures. The best performing models utilize larger neighborhoods, andthere is a clear trend that larger neighborhoods boost performance. Our proposedmethod improves results on all metrics by 1.35 - 3.14 % compared to a previouslydeveloped method for the same data using Gaussian scale-space features and a randomforest classifier. The porosities for the three HPC samples are estimated to 20.34,33.51, and 45.75 %, which is in close agreement with the expected porosities of 22,30, and 45 %. Interesting future work would be to let multiple experts segment thesame image to obtain more accurate ground truths, to investigate loss functions thatbetter correlate with the porosity, and to consider other neighborhood sizes. Ensemblelearning methods could potentially boost results even further, by utilizing multipleCNNs and/or other machine learning models together.

  HÄR KAN DU HÄMTA UPPSATSEN I FULLTEXT. (följ länken till nästa sida)