Hyperspectral imaging for the detection of malignant ovarian cancer metastases

Abstract

BACKGROUND: Ovarian cancer is the number eight cause of cancer-related mortality in women around the world and has an incidence of more than 295 000 and mortality of almost 158 000. The most important prognostic factor for the survival of advanced-stage ovarian cancer is the complete resection of all microscopic tumour tissue during cytoreductive surgery. Still, even after complete CRS and removal of all visible and palpable tumours, more than two-thirds of women with advanced-stage ovarian cancer experience recurrence due to microscopic residual tumour. Therefore, an intraoperative visualisation technique is of great importance to detect these microscopic tumours. In this pilot study, the feasibility of NIR hyperspectral imaging for the detection of malignant ovarian cancer metastases was evaluated ex vivo. METHOD: Patients with suspected primary ovarian cancer planned for cytoreductive surgery were enrolled in the study. Hyperspectral images from the resected tissue were acquired in the wavelength range of 665-975 nm. All hyperspectral data was preprocessed by image calibration, min-max normalisation and noise filtering. The the most important features were selected before implementing the hyperspectral data in the classifier. The linear SVM, RBF SVM, and a k-NN classifier were used to discriminate malignant ovarian tissue from the surrounding tissue. The performance of the classifier was evaluated by leave-one-out cross-validation. RESULTS: Ten patients were included in the study. The linear SVM classifiers had the highest performance with a sensitivity of 0.81, a specificity of 0.75, AUC of 0.83 and MCC of 0.41. Thereafter came the RBF SVM classifier followed by the k-NN classifier. CONCLUSION: This pilot study shows that hyperspectral imaging with the use of a linear SVM classifier is a promising technique to discriminate malignant ovarian tissue from the surrounding tissue. Hyperspectral imaging can scan a whole area, is fast, non-contact, non-invasive and can be used inside the operation room. However, before HSI can be implemented in real practice, further validation is required in vivo and technical enhancements need to be made such as enlarging the training data set, evaluate the SWIR wavelength range, include a registration algorithm which accounts for elastic deformation of the tissue, evaluate several feature selection methods, and compare other classifiers.