Multispectral imaging for intraoperative tumour detection in breast cancer surgery
File(s)
Author(s)
Shanthakumar, Dhurka
Type
Thesis
Abstract
On average 19% of patients undergoing breast conserving surgery (BCS) require re-excision due to positive margins. Augmenting a surgeon’s intraoperative visualization of margin disease could improve precision. Multispectral imaging (MSI) utilizes spectral differences between normal and tumour tissues to characterise pathology in real-time. The diagnostic accuracy of a custom-built MSI camera was assessed.
BCS patients were recruited to a single centre, prospective study (REC= 08/H0719/37). Multispectral images were acquired from each resection surface of freshly excised BCS specimens. Image pre-processing included data normalization and dimensionality reduction. Intraoperative radiography and gold standard histological analysis were used to extract ground truth. Four machine learning classifiers were used for data analysis.
46 patients were recruited. MSI images provide information about tissues up to depths of 3mm, therefore image analysis was conducted on resection surfaces where tumour was noted within this depth. 6 specimens were excluded due to gross contamination with blue dye or poor image quality. Therefore, 38 surface images from 23 specimens were included for final analysis. Logistic regression resulted in an area under the curve of 89% (SD ± 7), sensitivity 86% (SD ± 12), and specificity 80% (SD ± 8).
MSI can distinguish between normal and malignant breast cancer tissues. Unlike existing systems, MSI provides immediate visualisation and evaluates the entire resection surface. Future work will focus on adapting MSI to overcome spectral artifacts and improving image analysis using varying machine learning classifiers, as speed and accuracy will optimize surgical workflow.
BCS patients were recruited to a single centre, prospective study (REC= 08/H0719/37). Multispectral images were acquired from each resection surface of freshly excised BCS specimens. Image pre-processing included data normalization and dimensionality reduction. Intraoperative radiography and gold standard histological analysis were used to extract ground truth. Four machine learning classifiers were used for data analysis.
46 patients were recruited. MSI images provide information about tissues up to depths of 3mm, therefore image analysis was conducted on resection surfaces where tumour was noted within this depth. 6 specimens were excluded due to gross contamination with blue dye or poor image quality. Therefore, 38 surface images from 23 specimens were included for final analysis. Logistic regression resulted in an area under the curve of 89% (SD ± 7), sensitivity 86% (SD ± 12), and specificity 80% (SD ± 8).
MSI can distinguish between normal and malignant breast cancer tissues. Unlike existing systems, MSI provides immediate visualisation and evaluates the entire resection surface. Future work will focus on adapting MSI to overcome spectral artifacts and improving image analysis using varying machine learning classifiers, as speed and accuracy will optimize surgical workflow.
Version
Open Access
Date Issued
2025-02-02
Date Awarded
01/05/2025
License URL
Advisor
Darzi, Ara
Elson, Daniel
Leff, Daniel R
Publisher Department
Department of Surgery & Cancer
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)