JANUARY/FEBRUARY 2005 (Vol. 20, No. 1) pp. 5-9
1541-1672/05/$31.00 © 2005 IEEE

Published by the IEEE Computer Society
Danna Voth








Using AI to Detect Breast Cancer
Scientists are marshalling AI tools in the fight against the leading cause of cancer among women. Using digital information from traditional breast cancer detection techniques (see the sidebar ) as well as newer detection methods, new applications of data mining and neural networks are helping doctors detect cancers sooner. The AI tools can also reduce the number of biopsies required in the diagnostic process.
Image-enabled data mining
Working with new digital mammography technology, Chris Barnes, associate professor of electrical and computer engineering at Georgia Tech's Savannah campus, is using a new data mining technique that employs sensory data in queries. In particular, Barnes is working on image-enabled data mining software, based on the SOLDIER (source optimized lexicon digital expanded representations) query tool developed at Georgia Tech.
The tool uses pixel-driven queries to find information doctors can draw on when making diagnoses. It exploits a heterogeneous archive that holds both imagery and other patient data. Doctors use an image, such as a mammogram, to query the archive for relevant case histories.
The underlying query technique applies a successive-approximation code to both the patient's mammogram and archived imagery. This progressive data technique uses the patient mammogram to search a database for a somewhat similar image that also contains case history information. It continues refining the search to get a closer match, until it achieves the desired level of detail.
Figure 1 shows the user interface for the data mining software. "When part of the image provides a match within the compressed domain," Barnes says, "the refined versions of the code come into play to drill down into the archive to find similarity sets that are more and more like the part of the image used as a query."




Figure 1. Data mining software matches an image of calcifications from one mammogram to a database of case histories to find a similar pattern with potentially similar diagnoses.



The system analyzes the relevant imagery and associated point data to make an inference that can help the doctor make a decision in a clinical situation.
To use the system, a doctor points the cursor to an area on a mammogram, draws a box around it, and clicks a button. The computer then uses that region of the image as its query, searching the archive, and returning a dialog box containing case history information.
"The system does not require application-specific engineering to apply it," Barnes says. "It basically is just an interface tool between clinical image flow and archived image data."
Barnes says his system benefits the diagnostic process. "It can impact the speed of decision making, and it can lower the false-positive rate."
The system lacks a large archive of encoded imagery to draw on at this time. Barnes envisions a database built on the fly from clinical workflows at centers. Alternatively, a single institution or consortium might choose to build the historical archive.
Neural net training
Another scientist, also working with digital mammograms, is trying to help doctors better determine when a mass is malignant. Yulei Jiang, assistant professor of radiology at the University of Chicago, is developing a system that uses a perceptron neural network to analyze eight input nodes, converting the output node into a probability measure, which Jiang calls likelihood of malignancy.
"The advantage of artificial neural networks is that the solution is not restricted to linear form," Jiang says. The system is trained on a set of cases, using the leave-one-out method, which is a cross-validation technique for estimating generalization error based on resampling. A net is trained a number of times, each time leaving out one of the subsets. The omitted subset computes whatever error criterion is of interest. The system initially learns from digitized screen mammograms.
The eight input nodes represent features of calcifications, areas in breast tissue where tiny calcium deposits build up and might indicate the presence of cancer. Jiang's sys- tem looks at single calcifications as well as groups of calcifications, called clusters. The eight nodes are the shape-linearity measure of individual calcifications and the number of clusters, their sizes, shapes, average size, average size times contrast, uniformity in contrast, and uniformity in size times contrast.
The AI system feeds these nodes into the neural network to provide a statistical indication of the possibility that a group of calcifications is malignant.
The system was tested on 104 known cases from 10 radiologists and was measured in ROC (receiver operating characteristic) curves, a measure of the relationship between testing sensitivity and false-positive rates. "We've shown when we give the radiologist computer aid, they get a much higher curve," Jiang says. "We also show that they would have biopsied more cancers, and fewer patients with benign masses."
More recently, Jiang has tested the system on full-field digital mammograms and found similar results.
As Figure 2 shows, Jiang's system gives a doctor two views of the digitized mammogram—one a frontal top-to-bottom view called CC (cranio caudal), the other a lateral view. The computer helps to roughly point out areas of calcifications. The doctor examines the image to see the location and extent of the calcifications, then uses a mouse to draw a box around the calcifications for analysis. After the system gives its determination of the probability of malignancy, the doctor can then use the computer results to augment his or her professional opinion.




Figure 2. The radiologist marks the locations of calcifications with a white box on the image, and the computer estimates the likelihood of malignancy.



"Radiologists need to look at multiple features in the image to make a diagnosis," Jiang says. "I think the computer and mathematics have an advantage in that area."
AI-assisted infrared thermography
Infrared Sciences Corp. in Hauppage, New York, is developing a thermography-based cancer detection system that employs AI. In February 2004 the FDA approved the system as an adjunctive test for early detection of breast cancer.
The Breastscan IR system measures indications of breast cancer associated with angiogenesis, the formation of a blood supply to abnormal cells. The blood supply causes abnormal cells to respond differently from healthy cells to temperature changes; thermography measures such responses with infrared technology. ISC's system briefly blows cold air on the breast and then uses an infrared camera to record thermal images of the breast for analysis.
Thermography provides physiological breast data, as a PET (positron emission tomography—see the sidebar ) scan does, in contrast to the anatomical data that mammography or ultrasound provide. Thermographic data might be able to improve on PET scans, however, by indicating the possibility of cancer before an anatomical development grows large enough for other tests to detect.
Tom DiCicco, CEO of ISC, says the company added an AI component to its system to enhance its classification of equivocal or borderline test results. A cascading neural network with seven information inputs maps to a binary output. The seven inputs include such things as temperature differentials and asymmetries in the image.
The neural network was trained on equal amounts of biopsy-proving cancers and presumed normal cases, and has been retrained four times. "Our goal was to raise the number of people that pass this test without degrading the sensitivity of the test to cancer," DiCicco says. "Often those two things don't go together. In our case we were able to reach a compromise."
An oncology center, a radiology center, and a women's imaging center are currently using the system. The company recently raised money to give away more of the systems to get people familiar with and interested in using the technology. While ISC's thermographic system needs more testing to better determine the accuracy of its diagnoses, its potential for early indication of possible cancer development could encourage women to monitor for cancer more closely and engage in healthier activities.
It could offer a baseline test for young women. Unlike mammograms, the method is painless because it doesn't compress the breast—an appealing aspect to many women. However, early warnings for possible development of cancer can also create anxiety.
Neural networks and AI methods of pattern recognition can help indicate the presence of breast cancers sooner. By providing more detailed information from case histories, they might also reduce the number of biopsies doctors must perform. Armed with information earlier, patients might make better choices about actions affecting their health and enjoy greater success in fighting breast cancer.
PHONOLOGICA: QUANTUM PHYSICS MEETS AI?
Babies learn quickly to recognize their mothers' voices. Even a dog can easily recognize its master's voice. But computer applications based on voice recognition remain a challenging—some might say baffling—area of artificial intelligence.
Researchers at King's College, London, and Phonologica have been studying human speech models based on quantum physics principles—an investigation they hope will lead to advancements in the field.
"Speech recognition by computers suffers from some of the same limitations as other problems which the AI community has sought to address over the last few years," says Barbara Forbes, Phonologica's founder and the project's chief scientist. "Simply put, what humans can recognize almost instantaneously—people with very different accents, for example—computers either can't do, or take too long to do."
SCHRÖDINGER'S EQUATION
Forbes believes that the trouble with the current approach is that people in the field are throwing brute computer power at the problem without having an appropriate analog for how the human vocal tract manages to generate so much so quickly.
Enter quantum physics.
"The essence of our work is that we describe wave propagation within the vocal tract using an equation that has many similarities to the Schrödinger equation of quantum mechanics," Forbes explains. "This means we can use rather sophisticated mathematical tools of analysis that have been developed and tested in the field of quantum physics."
In the 1920s, Erwin Schrödinger proposed his famous equation, illustrating a central quantum physics tenet: One can't predict an event's actual movement, only the event's potential to occur. The concept turns out to have a surprising relevance to speech recognition.
"We have found many advantages in being able to talk about the shape of the vocal tract in terms of a potential function, rather than a simple area function," Forbes says. "The area function is just the cross-sectional area of the tract as a function of distance from the glottis, whereas the potential function is a higher-order function and is related to the curvature of the vocal tract.
"The physics of quantum mechanical potential functions is well known and is discussed in every textbook," she continues. "We've been able to borrow much of this and so can describe speech physics in a more simple but compact and powerful way than has previously been possible."
Forbes offers a striking example of the tangible benefits of applying quantum theory to the challenge of speech processing: Just six bits of information can encode around 27 vowels, the International Phonetic Alphabet's full vowel space. Similar results are on the horizon for consonants.
"There is still a lot that researchers don't understand about the physical mapping between the shape of the vocal tract and the audible speech sound," she says. "We want to define this mapping in a robust, compact way so that vocal-tract parameters can be identified from the speech sound and used in automatic speech recognition procedures."
Speech Production Versus Recognition
Forbes admits that this approach attempts to derive an alternative to the purely statistical methods that have dominated the field thus far.
"The current data-driven methods are not primarily structured around vocal tract physics, so 'adding in' speech production knowledge is cumbersome and computationally heavy," she says. "We intend to make the physical model the core of the 'engine' and optimize by statistical methods only at the next level of the architecture."
Timothy J. Hazen, a research scientist at MIT's Computer Science and Artificial Intelligence Laboratory/Spoken Language Systems Group, isn't convinced. Hazen believes that evaluations of Phonologica's research must consider the crucial difference between speech recognition and speech production.
"The main contribution of the Phonologica research is in the area of speech production, but understanding the physics of speech production may in fact be irrelevant to speech recognition," Hazen says. "For the specific problem of speech recognition, I'm skeptical that knowledge of the physics of speech production will be useful. For example, most humans have no knowledge about the physics of the vocal tract, and yet they have no difficulties recognizing speech."
Hazen points out that despite numerous prior attempts to incorporate speech production knowledge into speech recognition systems, none have yet to outperform current data-driven approaches.
"There have been, and continue to be, numerous attempts to incorporate knowledge of the speech production mechanism into speech recognition systems," he says. "Unfortunately, the problem of determining the dynamically changing configuration of the vocal tract purely from an acoustic signal is an extraordinarily hard problem that has stymied many researchers."
Taking It to Market
This problem hasn't stymied Phonologica's research team, however. What began as an academic pursuit—namely, Forbes' doctoral dissertation at King's College—has become a commercial venture. After obtaining venture capital funding to develop this research toward the marketplace, the university "spun out" the study to start-up company Phonologica in March 2004.
"Phonologica's work is at the stage where we believe it is extremely promising," says Grant Masom, Phonologica's chairman. "By going down a commercial, rather than purely academic route, we believe it will be possible to make the technology available to these developers in due time through partnerships and other funding arrangements where they can see a direct benefit rather than just funding purely academic work. Early days yet, but it looks promising so far."