| Three dimensional reconstruction and examination of microscopic images of tissue sections can enhance the study of normal and disease processes, particularly in studies concerned with structural changes. 3D histology volumes can also be used as ground truth in validating new volumetric medical imaging techniques. The accuracy of the histology volume reconstruction is hindered by sparse sampling of the clinical specimens, the deformation of the tissue during the histology processes, and variations in staining. The existing methods for histology volume construction of human breast specimens are either manual or applicable only to small cropped densely sampled regions.;In this thesis, we hypothesize that by using image registration algorithms, 3D histology volumes can be automatically reconstructed from stacked 2D images of histology sections. We present an image registration framework for histology volume reconstruction, which is also extendable for registration of the histology to ex vivo medical images. The framework is first validated using a pre-clinical model and then further improved by optimizing the deformable registration for histology images obtained from highly deformable human breast specimens. Finally, the histology volumes of five lumpectomy specimens are reconstructed using the proposed framework to study tumour features in 3D, specifically tumour volume and focality.;The proposed approach is shown to be suitable for 3D histology volume reconstruction from 2D histology images of breast lumpectomy specimens. The average target registration error for the breast specimens was measured as 0.47 +/- 0.31 mm based on manually selected landmarks and the optimized deformable registration was shown to significantly improve the registration comparing to rigid registration. It is also shown that this method has the potential to provide more accurate information about 3D tumour measurements, such as volume, and presence of multi-focality. |
