Training data selection for cancer detection in multispectral endoscopy images

Multispectral endoscopy images provide potential for early stage cancer detection. This paper considers this relatively novel imaging technique and presents a supervised method for cancer detection using such multispectral data. The data under consideration include different types of cancer. This poses a challenge for the detection as different cancer types may exhibit different spectral signatures. Consequently, it is not always feasible to transfer the knowledge learnt from one data set to another data set. In our approach, we select suitable training data for a given test set based on a similarity measurement between data sets. Experimental results demonstrate that the classification results can be significantly improved if a few data sets that are presumably similar to a given test set are selected for training instead of using all available data sets.