Chapter 2 Review Articles
2.3 Development of diagnosing cancer using hyperspectral image
2.3.1 Diagnosis of cancer from image
A fluorescent image was produced in the tissue that the light source irradiated. Because imaging provided two-dimensional information, we could easily mark the areas of lesions in the specific excitation wavelength. The fluorescence imaging techniques enabled real-time detection of cancer simple; it is inexpensive and has high sensitivity and specificity. Many studies have been performed using fluorescence imaging [6] to detect cancer in different organs, such as oral cancer and cancer of the lungs [7-12], the bladder [13], the colon [14-20] and the gastrointestinal tract [21-24].
Ina et al. used the laser-scanning fluorescence 351–364nm and 488nm excitation to detect cervical cancer [25]. Due to the essential dye, Mitotracker Orange, they could see the precancer’s cytoplasmic fluorescence in the bottom of the epithelium. They found that the intensity of fluorescence decreased with the development of cancer cells. In this study, only 10 patients were diagnosed, and the sample number was low.
Darren et al. observed the fluorescence images of oral lesions and normal tissues; the images were obtained from 56 patients and 11 normal volunteers [26]. They classified the images as normal and cancer in different fluorescence excitation wavelengths, 365, 380, 405 and 450nm, with the ratio of normalized red-to-green fluorescence and the autofluorescent image. The data were divided into two sets: a training set and a validation set. The training set included: 20%
invasive cancer, 28% dysplasia and 52% normal; the validation set included:
14% invasive cancer, 25% dysplasia and 61% normal. With 405nm excitation, the autofluorescent image showed obvious decreased intensity. It had the highest sensitivity (95.9%) and specificity (96.2%) in the training set, and 100%
sensitivity and 91.4% specificity in the validation set. They provided a noninvasive and sensitivity tool to diagnose oral cancer. They detected oral lesions using the decrease in autofluorescence images without observing the spectrum.
Catherine et al. used a hand-held device to evaluate oral cancer by location, fluorescence visualization (FV) status, histology and loss of heterozygosity (LOH) [27, 28]. First, they marked a blue line on the surface of the tumor diagnosed by the naked eye. After the light illuminated the tissues, the tissues provided direct visualization, and they marked a green line in the FV loss (FVL) area as the tumor margins. Lastly, they used LOH to analyze the FVL biopsies
from the tumor margins. In a total of 44 patients, the sensitivity was 98% and the specificity was 100%. However, they did not analyze the spectrum of biopsies.
2.3.2 Diagnosis of cancer from spectrum
Diagnosis based on spectrum has the potential to determine the change of material in the cancer cells. The spectrum was classified as the halogen spectrum and the fluorescence spectrum. When the halogen irradiated the section, we recorded the transmittance as the halogen spectrum. When the light source excited the tissue, the fluorescence spectrum was emitted by the tissue.
The spectral data were saved to a computer to be analyzed. Many studies have been performed using the spectrum method to distinguish between normal cells and cancer cells, and the method includes: Principal Components Analysis (PCA) [29-31], emission wavelength ratios [32-35], change of intensity [36-40]
and artificial neural networks [41, 42].
Irene et al. evaluated low-grade and high-grade dysplasia of Barrett’s esophagus (BE) by fluorescence, scattering properties, and enlargement and crowding of nuclei [43]. They tried to distinguish high-grade dysplasia from low-grade dysplastic and nondysplastic BE, and high-grade and low-grade dysplasia from nondysplastic BE. There were two peaks in the fluorescence with 337nm excitation; the decrease between the two peaks occurring in the 420nm was caused by the absorption of oxyhemoglobin; they combined the corresponding reflectance spectrum to compensate for the decrease.
At 337nm excitation, the line-shape of the spectrum shifted to the right during the progression from nondysplastic to low-grade, to high-grade dysplasia.
At 397nm excitation, the increase of intensity was found in the wavelength range 600–750nm. After the corresponding fitting of the reflectance spectrum, the scattering coefficient reflectance spectrum, μs, changed in different grades of dysplastic tissue. They also showed the enlargement of nuclei could be the characteristic, defining diameter > 10μm as enlargement of the nuclei. The analysis combining 3 techniques had a sensitivity of 93% and a specificity of 100%. Their analysis combined the spectrum and image, but the enlargement would be hard to define because of the irregular shape of the nuclei.
Hamed et al. detected cancer by using integral, support vector machines (SVM), spectral standard deviation and the normalized cancer index (NDCI) [44]. The halogen spectrum was normalized to calculate the reflectance using the following equation:
) wavelength range was 1000–2500nm, and they found the area under the spectral curve was higher in cancer than in normal tissues, while the slopes at 1200–1400nm were lower in cancer tissues. They also used SVM to classify tissue into normal and cancer tissues. They also compared the spectral standard deviation using the following equations, (2–2) in two dimensions and (2–3) in k1 and k2 were the range of wavelength bands, i and j were spatial coordinates, i1 and j1 were the area size of the predefined neighbor, C was a coefficient, R was the reflectance and Rav was the mean of reflectance. The NDCI was calculated using the following equation:
where NDCI was the normalized cancer index, C1 and C2 were coefficients, Rk was the normalized reflectance in wavelength k, d(Rk) was the derivative of Rk , k1, k2, k3, and k4 were the wavelength bands. Lastly, the specificity was 88%using integral, 80% using SVM, 82% using spectral standard deviation and 93%
using NDCI. However, they only analyzed the halogen spectrum without the fluorescence that could show the change of biochemistry.
Kevin et al. used the colonoscopy to detect colorectal cancer from two imaging modalities: visible and NIR autofluorescence imaging and hyperspectral reflectance imaging [45]. For normalization, they subtracted the background intensity and divided the corresponding brightfield images. In the autofluorescence imaging, they found that normal tissues emitted more autofluorescence than the cancer tissues in 515nm excitation did, but the normal tissues emitted less autofluorescence than the cancer tissues did in 567nm
excitation. Therefore, they divided the fluorescence intensity at the second peak with 567nm excitation by the fluorescence intensity at the first peak with 515nm excitation to diagnose the colorectal cancer, and the ratio was less than 1.96 in normal mucosa and more than 1.98 in cancer, as shown in Table 2-3. In the hyperspectral reflectance imaging, the result of a total of 7 samples (T1-T7) showed a great correlation between different stages of cancer except for tissue sample T5. However, the number of samples was too low to prove the method was successful.
Table 2-3: Tissue type, disease state, and 516/515nm ratio.
Sample Tissue type Disease state 567/515nm ratio
T1 Rextum Carcinoma 2.0003
T2 Sigmoid Colon Dysplasia adenoma 2.0655
T3 Colon Normal mucosa 1.9563
T4 Sigmoid Colon Dysplasia adenoma with possible cancer carcinoma in 404nm excitation [46]. The samples were all in hamsters, and the oral squamous cell carcinoma was induced by chemicals. With 404nm excitation, they found the intensity increased at 634 and 672nm peaks and decreased at 520 and 582nm peaks, as shown in Table 2-4. They said the decrease at 520nm was due to the reducing oxidized forms of riboflavin in tumor tissues; the increase at 630nm was due to the accumulation of porphyrin in the tumor tissues. After they defined the intensities of the peaks at 582, 634 and 672nm as A, B and C, respectively, they calculated A/B, B/C and A/C, as shown in Table 2-5. The value of A/B and A/C decreased during the progression from control to hyperplasia, to early cancer, to invasive cancer.
Table 2-4: Fluorescence intensities with 404nm excitation.
Peak A (582nm) Peak B (634nm) Peak C (672nm)
Control (n=6) 24.9±3.8 5.3±0.7 0.9±0.3
Hyperplasia (n=10) 21.7±4.7 6.4±2.4 1.2±0.3 Early cancer (n=5) 17.1±5.8 5.1±1.1 1.4±0.6 Invasive cancer (n=3) 13.0±9.0 27.0±19.1 8.1±4.7
Table 2-5: Ratio of fluorescence intensities with 404nm excitation.
A/B B/C A/C
Control (n=6) 4.7±0.5 6.2±1.6 8.0±9.4
Hyperplasia (n=10) 3.4±0.7 5.5±1.6 18.8±4.4 Early cancer (n=5) 3.4±1.0 4.2±1.4 13.6±4.5 Invasive cancer (n=3) 0.8±0.7 6.5±8.5 1.5±0.7
Brigitte et al. analyzed the fluorescence spectrum of colorectal cancer with 375–478nm excitation [47]. When they evaluated the spectrum from 478 to 700nm, they found that the ratio of the intensity in the 500–549nm to 657–700nm was a characteristic of the presence of cancer cells, and the critical value was 2.25. The sensitivity of this analysis was 97%, and its specificity was 95%. However, the number of the spectrum was less than 15 in one sample, too low to prove the method was successful.