University of Twente Student Theses


real-time tissue classification during vacuum-assisted biopsy procedures using diffuse reflectance spectroscopy

Lambregts, Maud (2022) real-time tissue classification during vacuum-assisted biopsy procedures using diffuse reflectance spectroscopy.

Full text not available from this repository.

Full Text Status:Access to this publication is restricted
Embargo date:7 October 2050
Abstract:Current limitations of a stereotactic Vacuum Assisted Biopsy (VAB) procedure of the breast are the underestimation of invasive carcinoma in biopsy samples, under- and oversampling of breast tissue, and it is hard to obtain microcalcifications for challenging locations. With quality control during the procedure, the diagnostic yield will increase, the number of biopsies will be reduced and the overall accuracy of the procedure will improve, resulting in reduced complications for the patient. The aim of this project is to give real-time feedback during a stereotactic breast biopsy procedure to better locate the VAB needle. An optical introducer is developed at the NKI-AVL to give more information about the VAB needle location with help of Diffuse Reflectance Spectroscopy (DRS). To provide real-time feedback with the help of that introducer, a tissue classification model needs to be developed to distinguish breast tissue during the procedure with help of DRS. Therefore, DRS measurements are acquired on breast specimens in an ex-vivo setting with the optical introducer. To track the DRS measurement locations, a top-view image of the breast specimen with projected DRS measurement locations is obtained before the DRS measurements. Labels with tissue types are required to develop a tissue classification model and distinguish breast tissue types. The ground truth tissue types are obtained by assessing microscopy sections (stained with hematoxylin and eosin (HE)) by the pathologist. During this pathology process, breast tissue is deforming due to e.g. fixation and waxing. To correlate the DRS measurements with the ground truth tissue types, a registration technique is needed to register the top-view specimen image and the HE image. This registration technique needs to correct for tissue deformations during the pathology process to create an accurate correlation. A deep learning-based multi-modal image registration model is developed to register HE images and top-view RGB images of the breast specimen. The model is trained with an existing dataset with those images. The registration model is based on VoxelMorph, a network architecture for image registration. The loss function is adapted to be able to register two images obtained with a different modality. Two types of learning are tested: supervised learning and unsupervised learning. For supervised learning, manual registered images are used for creating labels. Additionally, the data was augmented by creating different levels of artificial deformations on the images to mimic the tissue deformations. The best-performing deep learning-based registration model is trained in an unsupervised fashion (no ground truth labels are needed) and shows high evaluation scores (Dice of 0.97 and Mutual Information of 0.54 ). It shows a better Dice score compared to the manual approach, which means that the registration of HE and RGB images is improved. The best-performed model (according to new acquired data) is implemented in a pipeline to correlate the DRS measurements with the ground truth tissue types. The pipeline consists of multiple steps: data preparation, rigid registration of HE images with and without annotations, affine registration of RGB with and without projected measurement locations, deep learning-based registration of HE and RGB images, and label extraction. A breast tissue classification model is developed with the DRS data and the ground truth tissue type labels. The preliminary results of this model are promising: sensitivity of 95% and specificity of 98% to distinguish breast tumor tissue from healthy tissue. However, it requires more investigation to use the classification model for real-time feedback during a stereotactic biopsy procedure.
Item Type:Essay (Master)
Faculty:TNW: Science and Technology
Subject:44 medicine
Programme:Technical Medicine MSc (60033)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page