Discriminative Spectral Pattern Analysis for Positive Margin Detection of Prostate Cancer Specimens using Light Reflectance Spectroscopy

For localized prostate cancer, one treatment is prostatectomy which surgically removes the prostate gland. However, some undetectable cancer cells may be left as positive surgical margins, leading to a high risk of cancer recurrence. It is highly desirable to develop a portable and accurate classification methodology that detects positive margins on human prostate specimens immediately after their removal during surgery. This study applied data mining techniques on the light reflectance spectroscopy (LRS) data taken from ex-vivo human specimens and developed a novel classification algorithm that could enable real-time, positive-margin identification during surgery. Specifically, the LRS measurements taken from human prostate specimens ex vivo were classified to normal or cancerous tissue with support vector machines and were also classified to normal, cancerous and transition-to-cancer class with an ensemble of trees. The data in this study were highly overlapped and imbalanced among classes. We solved the overlapping issue by defining a middle class (transition-to-cancer), and by optimizing a moving spectral window through the range of LRS. To solve the imbalanced problem, we removed irregular tissue measurements, followed by application of random undersampling from the majority class. We achieved sensitivity and specificity of 100% and 82% for binary classification.