Pattern recognition for biomedical imaging and image-guided diagnosis

Pattern recognition techniques can potentially be used to quantitatively analyze a wide variety of biomedical images. A challenge in applying this methodology is that biomedical imaging uses many imaging modalities and subjects. Pattern recognition relies on numerical image descriptors (features) to describe image content. Thus, the application of pattern recognition to biomedical imaging requires the development of a wide variety of image features. In this study we compared the efficacy of different techniques for constructing large feature spaces. A two-stage method was employed where several types of derived images were used as inputs for a bank of feature extraction algorithms. Image pyramids, subband filters, and image transforms were used in the first-stage. The feature bank consisted of polynomial coefficients, textures, histograms and statistics as previously described [1]. The basis for comparing the performance of these feature sets was the biological imaging benchmark described in [2]. Our results show that a set of image transforms (Fourier, Wavelet, Chebyshev) performed significantly better than a set of image filters (image pyramids, sub-band filters, and spectral decompositions). The transform technique was used to analyze images of H&E-stained tissue biopsies from two cancers: lymphoma (three types of malignancies) and melanoma (benign, primary, and five secondary tumor sites). The overall classification accuracy for these cancer data sets was 97%.