University of Twente Student Theses


Intra-operative assessment of resection margins by three-dimensional ultrasound in patients with squamous cell carcinoma

Bekedam, N.M. (2020) Intra-operative assessment of resection margins by three-dimensional ultrasound in patients with squamous cell carcinoma.

[img] PDF
Abstract:Surgical excision is the most common treatment for tongue squamous cell carcinomas (TSCC). Surgeons aim to remove the tumor with a minimal resection margin of 5 mm to reduce the chance of recurrence. Currently, there is no intra-operative assessment available to determine if a 5 mm resection margin is achieved. To improve the surgical precision of TSCC resections, this research aims to provide surgical guidance during resections of TSCC using three-dimensional (3D) ultrasound (US) to minimize close resection margins. This research was divided into three parts which together investigated the feasibility of 3D US for intraoperative assessment of surgical resection margins of tongue squamous cell carcinomas. The first part of this research provides a better understanding of data acquisition and reconstruction of 3D US. In a phantom study, the influence of the 1) reconstruction algorithm, 2) sweeping method, 3) US transducer frequency, 4) stabilization rails and 5) observer was investigated. The accuracy of the 3D US volumes was evaluated by the signal-to-noise ratio (SNR), contrast-to-noise ratio (CNR), derivative along a scan line and the Full Width at Half Maximum (FWHM) of the peak of the derivative of the pixel intensity. The results show that data acquisition was performed best using the highest US frequency possible, a single sweep method assisted by rails and performed by a single operator. This study could not identify a reconstruction algorithm performing better than others. The second part of this research investigated as a proof of concept that deep learning is a feasible technique for fast automatic multi-class segmentation of tongue specimen and tumor in 3D freehand US volumes. The multi-class segmentation of tongue specimen and tumor was split into two binary segmentation problems by adopting the cascade strategy. Two identical UNet models were trained upon their own dataset (from a total dataset of 44 3D US volumes of 8 patients) and the influence of the loss function (Dice and binary cross-entropy (BCE)) and data augmentation was investigated. Evaluation based on the Dice similarity coefficient (DSC), showed 86% DSC (BCE loss with data augmentation applied) predicting the specimen and 18% DSC (Dice loss and data augmentation applied) predicting the tumor. The third part of this research explored the correlation between resection margins assessed by 3D US and histopathology. This study included 8 patients of which the resection margins of TSCC were assessed intra-operatively by 3D US and post-operatively by histopathology. The correlation between the measurements by 3D US and histopathology was computed by the Pearson correlation coefficient. The results showed that the measurements of resection margins by 3D US and histopathology do not correlate statistically significant, meaning that 3D US could not provide correct intra-operative feedback to the surgeon. Future research should focus on expanding the dataset and improving the data acquisition, by utilizing a high frequency US transducer and stabilization rails. In addition, remodeling of histopathological slices into 3D models and registering those 3D models towards the 3D US models could help the radiologist annotating more accurately. Also, research should investigate which hyperparameters in the deep learning models perform superior to obtain maximum DSC in predicting the specimen and tumor in 3D US volumes. Eventually after all these improvements, it is speculated that recalculating the correlation between the resection margins measured by 3D US and histopathology in tongue tumor specimens could be statistically significant.
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