Hand-held scanner and software spot melanoma

Dermoscopic image-analysis system screens patients for potentially dangerous lesions.

Oct 1st, 2004
Th 160598

Dermoscopic image-analysis system screens patients for potentially dangerous lesions.

By C. G. Masi, Contributing Editor

Melanomas-cancerous lesions in the pigment-bearing basal layers of the skin-appear as small dark-brown or black spots with sharply defined, but usually irregularly shaped, edges. The spots contain slightly contrasting darker spots within them.


Ana Vijuk, technical officer at Polartechnics, demonstrates SolarScan scanner hand-held dermoscope with a three-CCD color imager to Australian Science Minister Peter McGauran.
Click here to enlarge image

Melanoma is the most deadly form of skin cancer, yet it is also the most treatable, with a cure rate for early-stage melanoma of almost 100%. The problem is identifying the small percentage of skin lesions that develop into melanoma soon enough to treat them effectively. If not found early, the survival rate drops dramatically. If the lesion grows to a depth of more than 3 mm, the survival rate drops to only 59%.

Unfortunately, accurately diagnosing this disease is difficult. Many benign, or even natural, skin spots or blemishes superficially resemble melanoma, and most general practitioners see only a handful of melanomas in their career. This leads to many false-positive diagnoses by physicians.

Australia has the highest incidence of skin cancer in the world because of its geographical location and the fact that most of its inhabitants have fair skins and enjoy an outdoor lifestyle. Most clinicians in Australia aggressively attack any suspect skin lesion. Consequently, Australians perform almost three-quarters of a million skin-cancer-removal operations-many of them unnecessary-each year, costing more than $300 million to the federal government. Other developed countries have a similar experience.

Recognizing this high false-positive rate, many patients ignore potentially dangerous lesions (effectively creating false negatives) until it is too late. Thus, despite physicians’ best efforts, melanoma claims the lives of more than 1000 people in Australia each year, 7700 people in the USA, and 1600 people in the UK. The rate of melanoma cases worldwide is growing at the rate of 6% per year-faster than the rate of any other cancer. In addition, since 1973 the mortality for melanoma has increased by 50%.

DERMASCOPIC DIAGNOSIS

An accurate, reliable, and inexpensive system to diagnose melanoma is needed. The most promising modality is epiluminescent microscopy (ELM, or dermoscopy). ELM illuminates the pigmented layer of the skin with white light while eliminating reflections from the horny outer layer (epidermis) by trapping a layer of clear oil between the skin and a glass plate (see Fig. 1). A ring source sends light around the optical path of a low-power microscope so that it enters the skin at a wide angle. Reflections from the upper glass surface are specular and aimed away from the magnifier’s entrance pupil. The oil trapped between the glass lower surface and the skin suppresses diffuse reflection from the rough epidermis by matching its refractive index.


FIGURE 1. Dermoscope uses a flat plate and oil film to prevent reflections from the epidermis from entering the imaging optics. The oil film provides a refractive index match that eliminates diffuse reflection. Specular reflection from the glass window�s top surface misses the optical system�s entrance pupil. A short-focus macro lens images the lesion under scrutiny onto the Toshiba IK-TU61 CCD camera sensors.
Click here to enlarge image

Researchers at Polartechnics and The Sydney Melanoma Unit (SMU) of the Royal Prince Albert Hospital conceived of solving the clinician-inexperience problem by mating a dermoscope with a machine-vision system running software capable of discriminating between benign spots and early-stage melanomas with high accuracy. SMU provided the diagnostic expertise, having the world’s largest database of melanoma patients (more than 15,000 entries) complete with dermoscopic images correlated with treatment outcomes. Polartechics provided the system-engineering and manufacturing capability and asked the Commonwealth Scientific & Industrial Research Organization to provide the image-processing expertise (see photo on p. xx).

The front end of the resulting SolarScan system consists of a hand-held scanner unit containing a dermoscope optical system that focuses light (piped from an LED via optical fiber to form a ring source) onto a Toshiba IK-TU61 three-CCD color minicamera with LVDS output built into the scanning head. Each CCD sensor is 760 ¥ 560 pixels. A glass window closes the unit’s front end and serves as the dermoscope’s glass plate. A FlashBus MV Pro PCI frame grabber from Integral Technologies in a Dell PC running Microsoft Windows 2000 captures a scanner image frame and forwards it to the image-analysis software (see Fig. 2).


FIGURE 2. Image data are transferred from hand-held scanner head to a frame grabber plugged into a PCI slot. Image-analysis software removes artifacts from the image file, then extracts 80 quantitative diagnostic characteristics from the image and statistically compares these characteristics to similar ones derived from the SMU database.
Click here to enlarge image

The scanner also uses four color swatches as calibration targets at the field-of-view corners to provide an accurate color reference for the image-processing system. The image-processing system later masks these swatches out of the image, as well as finding and removing images of hairs and any remaining trapped air bubbles. In use, the clinician applies a layer of clear oil such as mineral oil to the patient’s skin over the suspect lesion. The clinician then lightly presses the scanner over the spot to exclude air bubbles trapped between the glass and the patient’s epidermis.

ANALYSIS SOFTWARE

The software has two components: an image-analysis pipeline and a diagnosis model. The image-analysis pipeline isolates the lesion from extraneous parts of the image and extracts features that provide diagnostic information. It first calibrates the image color and brightness using the four color swatches, removing session-to-session and instrument-to-instrument variability and comparing information it extracts from the patient’s image to that from the SMU database images.

Next, the software removes artifacts present in the image that have no diagnostic value. It detects bubbles by looking for local variations in image brightness that have a sharp edge. Detecting hairs is more complex. The system looks for the intersection of a series of algebraic closings by linear structuring elements.

Finally, the system knows where the calibration artifacts are located in the image. The diagnostic system relies on image color as a significant component of the diagnostic information. It looks at both absolute colors and relative colors within the image. While relative color might be calculated without reference to a standard, the standard swatches are needed to calibrate the image colors absolutely. Once this calibration information has been extracted, the swatches are no longer needed and can be excised from the image.

The next step is to separate the lesion from the surrounding unaffected skin. At first, the software attempts to make this determination on its own using a seeded region growing (SRG) technique. The operator has to decide whether to accept the result. If the operator deems the automated system’s outline for the lesion unsatisfactory, he or she has the option of switching to a semiautomated algorithm (see Fig. 3).


FIGURE 3. System identifies melanomas by progressively extracting information from the image. From the raw image of a melanoma (left), a yellow line is added (center) to isolate the melanoma’s edge. The area enclosed and the irregular texture of the border help indicate that the lesion is cancerous. Measures of pixel darkness, color variation, and other characteristics (highlighted by false colors in the right-hand image) complete the diagnosis.
Click here to enlarge image

The main difficulty with an SRG technique is finding seeds to start. The automated system begins with a derived pixel quantity (P1) that is roughly equivalent to lightness. Pixels with P1 within the lowest 20% are seeded as part of the lesion. The lightest 20% are seeded as showing skin. The software then looks at each pixel’s position within a pixel-brightness histogram. Generally, the histogram has two main peaks-one representing lightness variation within skin regions and the other representing variations within the lesion. The SRG method assigns pixels to the lesion or skin regions according to where they fall in this bimodal histogram.

While the histogram is, in the main, bimodal, it typically shows additional peaks especially within the lesion. The SRG algorithm identifies these smaller clusters, providing an ordered series of image segmentations from darkest to lightest.

One of the most significant variables in subsequent statistical analysis is color, with melanomas being more colorful (larger scatter over the RGB color space) than benign lesions. The developers have found that both absolute and relative colors are diagnostically useful.

Texture has also proved to be a necessary diagnostic component. Some melanomas proved not very colorful but could be distinguished from benign lesions on the basis of texture. Melanomas tend to have less regular textures and less symmetry.

Altogether, the researchers ended up with a list of 100 characteristic features that, taken together, could distinguish melanomas from benign lesions with a false-negative rate below 8% and a false-positive rate below 38%. That is, it caught more than 92% of melanomas and correctly identified better than 62% of benign lesions. These results are comparable to the diagnostic ability of experienced skin specialists and superior to that of most general practitioners.


Early detection saves lives

Click here to enlarge image

“The best way to treat a melanoma is to catch it early. Early detection saves lives. Unfortunately, the diagnostic accuracy in the best of hands is only 60%, so we try to use other methods to help us take better care of our patients,” says Dr. Harold S. Rabinovitz at Skin & Cancer Associates (Plantation, FL, USA; www.drharold rabinovitz.dermdex. net).

“The SolarScan will aid the physician by being an independent aid to make a diagnosis of melanoma. It is currently undergoing a cooperative effort by groups in the United States, Australia, and other areas to train it to separate melanomas from atypical moles. I believe the machine will be a better diagnostician than the physician on a specific individual lesion. The combination of an expert physician and a computer-aided diagnosis will improve the quality of medicine.”


Company Info

Commonwealth Scientific & Industrial Research Organization Dickson, ACT, Australia www.csiro.au
Integral Technologies Indianapolis, IN, USA www.integraltech.com
Polartechnics Sydney, NSW, Australia www.polartechnics.com.au
The Melanoma Foundation Sydney, NSW, Australia www.medicine.usyd.edu.au/melanoma
Toshiba Irvine, CA, USA www.cameras.toshiba.com

More in Cameras & Accessories