Abstract
A hybrid neural network (HNN) developed by Physical Optics Corporation (Torrance, CA) is helping a team of scientists with the San Diego Veterans Administration Medical Center and University of California, San Diego Pathology Department automate the detection and identification of Tuberculosis and other mycobacterial infections.
Processing data collected on a Finnigan 4500 gas chromatography-mass spectrometer, the researchers are relying on the HNN's intricate parallel processing scheme to mimic some of the pattern recognition and deductive reasoning capabilities of the human brain.
In a recently concluded test, the SSMART (spectral signature matching for automatic recognition and testing) hybrid neural network software successfully identified mycobacterial infections targeted in the study within one second. Scientists are hopeful that this experimental automated technique will help accelerate TB detection to the point where it can be accurately diagnosed in less than one week, compared to the four to six weeks currently required with conventional techniques.
Neural networks consist of highly interconnected layers of electronic elements called artificial neurons or nodes and have unprecedented capabilities for on-line, real-time interpretation and evaluation of spectroscopic data (see Figure 1). The essence of neural networks is to embed learning and non-linear processing algorithms into parallel interconnected schemes that lead to faster and more robust computations than are routinely possible with conventional signal processing methods. HNN systems can be customized for OEM customers or, as in the case of the tuberculosis research group, for end-users wishing to automate existing instruments or processes.
GC/MS spectrum showing selected ion monitoring (SIM) from a 21-day culture of M. tuberculosis sample.
In the neural network used by the research team, the neurons are organized into four neuron layers: one input, two hidden and one output. The HNN contains 1400 input neurons in the first layer, 40 hidden neurons in the second layer, 10 hidden neurons in the third layer and 10 output neurons in the fourth layer. Each of these neurons is connected to every neuron in the prior layer and every neuron in the succeeding layer (see figure 2), All connections have a weighted value, much like a valve or programmable circuit array that attenuates or magnifies transferred signals.
Schematic of the neural network concept.
Before a neural network can work on a specific task it must be trained to recognize critical characteristics of the input data. In the case of the TB research, GC/MS data were digitized and sent to a 50 MHz 486 PC for training and analysis, a process that took less than 15 minutes. From this extracted information, the neural network learned all the possible spectral variations, system calibration errors, and most importantly, the behavior of the system noise. The 10 output neurons represent 10 different species of mycobacteria, including TB and avium, among others.
Each output neuron gives an output value normalized between 0 and 100 to indicate the confidence level of the classification result (see Table 1).
Neural Network Output for Mycobacteria Analysis (confidence level 0 to 100)
Like all mathematical systems, the SSMART system works best when it's fed appropriate, detailed information, often including noisy signals, rather than just indiscriminate raw data.