Recognizing the terahertz response associated with skin cancer

The difficulties with using THz as a diagnostic tool lie not so much in obtaining data but in extracting useful information. Changes in the THz pulse that are caused by the sample can be difficult to distinguish from those caused by long-term fluctuations in the driving laser source. Increased contrast in THz images can differentiate between normal and cancerous tissue.^1–3 However, this increase has been attributed to greater water content of the diseased tissue,¹ which occurs in many diseases of the skin.

The spectroscopic responses of tissue samples may hold vast information, since many biomolecules have resonant peaks in the THz range due to phonon and intermolecular vibrations. However, sophisticated computational pattern recognition and related multivariate statistical tools are needed to analyze and classify these data.

Pattern classification⁴ is a technique for assigning samples, such as terahertz data, to one of several classes, such as cancerous and non-cancerous. First, the data to be classified is preprocessed. Next, features are extracted for each data set, and a vector consisting of the strength of each feature is passed to a classifier. A training set, consisting of data vectors whose class is known, is used to create a model. The model is then tested on a different test set, and the classes it predicts are compared with known classes of data in the training set.

Feature extraction and classification are usually two separate problems, each with its own difficulties. There are many algorithmic approaches to classification, which differ greatly in terms of their complexity and sophistication.

We have explored a highly sophisticated technique for classification, using support vector machines (SVMs),^5,6 on a simple data set⁷ consisting of THz spectral data collected from osteoblast (normal bone tissue) cells and osteosarcoma (a type of bone cancer) cells. We preprocessed the raw data using finite impulse response (FIR) filter coefficients⁸ that highlight subtle differences between cells that are normal and those that are cancerous.

The goal of SVM learning is to find the boundary, in the vector space of the extracted features, that best separates one class of samples from another. The vectors near the boundary are the ‘support vectors’ that are used to construct the boundary, which is then used as a classifier function. In most cases, a plane in feature space does poorly at separating the vectors from different classes, but a non-linear surface does better. The beauty of SVMs is that such complex boundaries can be constructed by simply modifying a kernel function.

Using the FIR filter coefficients as input, the SVMs were able to distinguish the normal and cancerous cells with an accuracy of 70.2%, which is reasonable given the extremely small sample size. This figure is substantially better than visual screening but is much lower than standard histochemical diagnosis. A larger data set than the one used in these preliminary studies would lead to a more accurate classifier which could answer more specific questions, such as type of cancer or developmental stage. Preprocessing would also improve accuracy. Other members of our group have been working on techniques for removing noise from THz data that significantly improves classification^9,10.

These results show that, in principle, THz spectral data, collected in a noninvasive way, could be automatically analyzed using a classifer to determine whether or not a sample is cancerous. Interestingly, the features used to do this are not necessarily physical. Current screening, which is done by visual observation, is based on specific physical features that can be quite subtle. Perhaps a classifier can tell us something new about what features and feature combinations to look for in visual screening.