A well-known legend has it that one of the greatest scientists and inventors of antiquity, Archimedes of Syracuse, stepped into a bath only to eject and propel himself naked throughout the city, yelling “Eureka!”, Greek for “I have found (it),” thus celebrating his discovery of how to measure the volume of irregular objects. Whether this indeed happened or not remains an open question, but a few important lessons can certainly be learned from this story. The first lesson is that good ideas could occur to us while we are taking a bath or a shower. The second lesson is that scientific or technological problems often seem to be difficult before a brilliantly simple solution is found. Does this second lesson hold true today? This is the question we will try to answer using an example of our own journey into discovery.

altWhile we do not have to solve Archimedes’ problem again, we still have problems to solve and discoveries to make. This is a brief story of our contribution— MelaFind®—the world’s first automated noninvasive instrument to help detect the deadliest form of skin cancer (cutaneous melanoma) at its most curable stages. (See Figure 1)

Challenge and Opportunity

MelaFind was developed by a group of research scientists who were working on defense research projects. After a meeting with melanoma expert Dr. Alfred Kopf, the group was inspired to start working on detection of cutaneous melanoma. The goal was to develop a device that would save lives and make a positive impact on humanity. Although it took significantly longer to accomplish this goal than we envisioned, it is rather remarkable that even today there are no other solutions to this problem that were validated in rigorous blinded prospective clinical trials and approved as an automated medical device with a proven diagnostic profile.

The fundamental difficulty of this particular problem is that unlike many diseases that humans can objectively diagnose, or unlike many military targets that our group was working with, melanomas in their most curable stages often lack any known and specific physical or chemical markers. The empirical characteristics, such as clinical A-B-C-D for Asymmetry, Border Irregularity, Color Variegation, and Diameter > 6mm, are also present in many benign skin lesions. An example of asymmetry is shown in Figure 2. Unfortunately, the prognosis for advanced melanoma is not good.

A New Approach

We decided to develop a machine that can learn and make discoveries based on a large number of histologically confirmed examples, both malignant and benign. While the field of machine intelligence is not new, I think our project represents one example in the next chapter of research; specifically, we, as an intelligent species, seem to be reaching the limit of making discoveries by virtue of the sheer “Eureka!” moment or manual analysis of the data, because most of the “simple” problems seem to have already been solved. By “simple,” of course, we do not mean that these problems did not appear to be difficult before they were solved, rather that their solutions turned out to be understandable, as was the brilliant discovery by the naked man from Syracuse.

In the future though, we will have to rely more and more on a mutually complementary partnership between human and machine intelligence to analyze the available wealth of information, to arrive at solutions to today’s tough problems. We believe that humans will remain designers of “machine inventors,” but the search for the discoveries themselves will be delegated to increasingly intelligent computers.

The most amazing consequence of this new approach, as our own experience with MelaFind has demonstrated, is that the discoveries made by machines, the very working recipes to the problems we seek answers to, may be so complex that they will not be completely understandable to humans, the creators of these “electronic Archimedeses” of the future. Just as we do not completely comprehend all the details about how our brain works, we will have to accept the humbling fact that even our own creations may produce answers beyond our understanding.

I have spent 14 years working on MelaFind and some of my more gifted colleagues have spent even more, and yet, in all honesty, I can’t explain exactly how it works. Some of its design follows the principles employed in machine (and often biological) vision. Specifically, the device has a chain of algorithms that, starting from the multispectral (from blue to near infrared) images, arrive at an evaluation of a lesion. The chain is as follows: image calibration, segmentation, automated image quality control, feature extraction, and classification.

We begin with image calibration. The purpose here is twofold: to remove or reduce the systematic components of the “noise” and to ensure that lesion information does not depend on the instrument used to acquire images.

Segmentation then automatically locates the lesion and determines its boundaries.

Figures 2a and b - Above, in this classic asymmetrical melanoma, the left side of the lesion is much thicker than the right side. Below, a normal mole is shown. (Credit: National Cancer Institute)
Next is automated image quality control. At this stage, any foreseeable imaging problems (like hair, air bubbles trapped in interface liquid, movement of the instrument, etc.) must be detected and, if their presence is found to be excessive or significantly interfering with the lesion information, the machine gives a recommendation on how to correct the problem.

Feature extraction then converts the segmented portion of the image (corresponding to a lesion) into a set of characterizing values (called features) that are meaningful for classification purposes. For example, the A-B-C-D rule is an example of four features currently employed by dermatologists to identify melanomas. Most of the features that MelaFind uses are far more complex than simply A-B-C-D.

Classification combines several extracted features in multidimensional space in order to make a decision on whether the lesion presents with a high or low degree of morphological disorganization. MelaFind operates in 75-dimensional space to arrive at this decision.

This final multi-dimensional classifier recipe, as well as the choice of 75 particular features out of a pool of more than 1,000 candidates, is a product of machine training and its composition makes little sense to us, the developers of the machine that trained MelaFind.

Ironically, the phrase that Archimedes was shouting after his famous submersion into the body of water, the very phrase that became one of the symbols of serendipity, also gave birth to the entire scientific discipline called heuristics—the set of techniques of problem solving based on prior confirmed experience, not the systematic analysis of the problem. MelaFind is one example of a heuristic approach to the problem.