A method of semi- automated ident- ification of biomaterials, now under development, is based on the use of a combination of neural-network and of self- organizing, molecular, and supervised-learning algorithms to analyze multispectral images. The method is especially applicable to such biomaterials as tissues and cells depicted in multispectral images of biopsy specimens. A basic premise of this method is that tissues and cells that are or may be malignant can be distinguished from healthy cells by morphological and molecular differences that manifest themselves as observable differences in multispectral images.

Images of a Specimen are acquired in multiple spectral bands, then digitized, then processed by neural-network and other algorithms to identify portions that merit scrutiny by a human expert.
Heretofore, it has been necessary for a pathologist to look at all the cells in a biopsy specimen under a microscope in order to determine the degree of malignancy. This exam- ination process is fatiguing, and most cells in a typical specimen are not malignant; as a result, the process is time-consuming and usually results in many errors. An automation method like this one could reduce examination time and fatigue while increasing the accuracy of diagnosis by quickly analyzing an entire specimen and identifying those portions of the specimen to which the pathologist’s attention and expertise could be applied most productively.

The figure depicts the major steps of the analysis process according to this method. A specimen is first placed on a microscope slide. Under an epifluor- escence microscope, the specimen is exposed to light in a wavelength band selected by means of a liquid-crystal tunable filter. At present, transmission images of a specimen in approximately fifty 5-nm-wide wavelength bands that span the range from 470 to 710 nm. (In future versions, the wavelength range may be divided into a greater number of narrower bands.) The image in each wavelength band is detected and digitized (at present, to 12 bits for each pixel) by a monochrome digitizing electronic camera. The resulting multispectral image data can be represented as a stack of the images in the various wavelength bands: in the art of multispectral imaging, such a stack, or the data that it represents, is denoted an image cube.

For the purpose of identifying portions of an image that merit a pathologist’s attention to determine the degree of malignancy, an image cube is processed by a number of algorithms, some of which implement artificial neural networks:

  • The image cube is mathematically filtered to extract salient features, including sizes, shapes, and orientations of cell nuclei.
  • The unsupervised-self-organizing attribute of neural networks is exploited to partition the image cube into classes based on similarities among spectral bands in the pixels.
  • A combination of the self-organizing and supervised-learning attributes of neural networks is exploited to enable the use of a complex database of image cubes of known benign and malignant cells as a guide to further classification of unknown cells.

Development efforts have included the creation of a database of image cubes of known benign and cancerous human prostate cells. In tests thus far, the method as applied to image cubes of unknown prostate cells yielded correct classification in 98 percent of the cases.

This work was done by Hamid S. Kohen of Caltech for NASA’s Jet Propulsion Laboratory.