Classification algorithm for tumor detection in hyperspectral images.
- Better classification algorithms than currently available for hyperspectral images.
- Improved segmentation of diseased from healthy tissue and improved sensitivity, specificity, and accuracy compared to the gold standard.
Surgical resection of tumors is a common treatment option for cancer. There is a need for verification of complete removal of cancerous tissue. Unfortunately current diagnostic imaging techniques require cumbersome equipment that cannot be easily accessible during surgery. Hyperspectral imaging may provide a non-invasive detection method for marginal cancerous tissue left behind following the removal of the tumor; and it can be performed using equipment of a much smaller size and greater mobility than the current gold standard.
A hyperspectral image consists of a three-dimensional array representing a two-dimensional spatial image. Previous research has demonstrated the feasibility of support vector machines (SVMs) to perform pixel-wise classification of hyperspectral images. Emory researchers have expanded on these techniques and introduced a spatial component to the classification methods. They have developed a method that relies on SVM classification and probability data to determine which pixels are most likely to be classified as cancerous or healthy tissue. These pixels are then determined as markers for which spanning trees are grown. These trees are minimized within a minimum spanning forest to segment the hyperspectral image. This method increases the accuracy of the SVM by incorporating spatial information within the spanning forests along with the spectral information the SVM uses for classification. Previous algorithms calculated weightings for the eight nearest neighbors of a pixel. The algorithm described here calculates the dissimilarity weighting between the eight neighbors and also includes the dissimilarity between the adjacent pixel and of the region that contains the pixel itself.
Algorithm has been developed and shows enhanced sensitivity, specificity, and accuracy.
View our feature innovation