Band-Selection of a Portal LED-Induced Autofluorescence Multispectral Imager to Improve Oral Cancer Detection

Abstract

This aim of this study was to find effective spectral bands for the early detection of oral cancer. The spectral images in different bands were acquired using a self-made portable light-emitting diode (LED)-induced autofluorescence multispectral imager equipped with 365 and 405 nm excitation LEDs, emission filters with center wavelengths of 470, 505, 525, 532, 550, 595, 632, 635, and 695 nm, and a color image sensor. The spectral images of 218 healthy points in 62 healthy participants and 218 tumor points in 62 patients were collected in the ex vivo trials at China Medical University Hospital. These ex vivo trials were similar to in vivo because the spectral images of anatomical specimens were immediately acquired after the on-site tumor resection. The spectral images associated with red, blue, and green filters correlated with and without nine emission filters were quantized by four computing method, including summated intensity, the highest number of the intensity level, entropy, and fractional dimension. The combination of four computing methods, two excitation light sources with two intensities, and 30 spectral bands in three experiments formed 264 classifiers. The quantized data in each classifier was divided into two groups: one was the training group optimizing the threshold of the quantized data, and the other was validating group tested under this optimized threshold. The sensitivity, specificity, and accuracy of each classifier were derived from these tests. To identify the influential spectral bands based on the area under the region and the testing results, a single-layer network learning process was used. This was compared to conventional rules-based approaches to show its superior and faster performance. Consequently, four emission filters with the center wavelengths of 470, 505, 532, and 550 nm were selected by an AI-based method and verified using a rule-based approach. The sensitivities of six classifiers using these emission filters were more significant than 90%. The average sensitivity of these was about 96.15%, the average specificity was approximately 69.55%, and the average accuracy was about 82.85%.

Keywords: AI-based band selection; LED induced autofluorescence; multispectral imager; oral squamous cell carcinoma; rule-based band selection.

Figure 1

Mechanical drawing and explosion drawing of the LIAF multispectral imager.

Figure 2

Composition of the 4CH LIAF multi-spectral imager and 8CH LIAF multi-spectral imager, including excitation lights, LED current, band-pass filters, a long-pass filter, and a CMOS sensor.

Figure 3

Flowchart of the trials.

Figure 4

(a)Tumor regions and normal regions, which are captured are marked as a white circle. (b) LIAF captures the tumor or normal points. (c) Captured image with a circle indicating the opening scope of the probe.

Figure 5

(a) Gaussian distribution of normal and tumor data. The Gaussian distributions of normal and tumor data are drawn as a solid line and dotted line, respectively. The crossing point marked as “◆” represents the optimized threshold. (b) The ROC curve of normal and tumor data in one classifier. The crossing point marked as “•” represents the optimized threshold.

Figure 6

Spectral radiant flux of the excitation LEDs.

Figure 7

Spectral transmittance of the emission filters and RGB color filters of the 4CH LIAF multi-spectral imager [43,44,45,46].

Figure 8

Spectral transmittance of the emission filters, long-pass filter, and RGB color filters of the 8CH LIAF multi-spectral imager [43,47,48,49,50,51,52,53].

Figure 9

The AUC statistic of the classifiers with 30 spectral bands in Exp_1, Exp_2, and Exp_3. The data in each classifier was used to calculate the AUC (A + B). The solid red points are the average AUCs. The highest and lowest black bold dash lines are the maximum and minimum AUC. The highest and the lowest blue dash lines are third and first quartile AUC. The mid-blue dashed lines are the median AUC.

Figure 10

The AUC statistic of the classifiers with 30 spectral bands in Exp_1, Exp_2, and Exp_3. The data in group A was used to calculate the AUC. The solid red points are the average AUCs. The highest and lowest black bold dash lines are the maximum and minimum AUC. The highest and the lowest blue dash lines are third and first quartile AUC. The mid-blue dashed lines are the median AUC.

Figure 11

Statistic AUC of the classifiers with 30 spectral bands in Exp_1, Exp_2, and Exp_3. The data in group B was used to calculate the AUC. The solid red points are the average AUCs. The highest and lowest black bold dash lines are the maximum and minimum AUC. The highest and the lowest blue dash lines are third and first quartile AUC. The mid-blue dashed lines are the median AUC.

Figure 12

(a) Sensitivity, specificity, and accuracy of the classifiers using the 505 nm emission filters correlated with the green filter. (b) Sensitivity, specificity, and accuracy of the classifiers using the green filter. (c) Sensitivity, specificity, and accuracy of the classifiers using the 470 nm emission filters correlated with the blue filter. (d) Sensitivity, specificity, and accuracy of the classifiers using the 550 nm emission filters correlated with the green filter. (e) Sensitivity, specificity, and accuracy of the classifiers using the 532 nm emission filters correlated with the green filter. (f) Sensitivity, specificity, and accuracy of the classifiers using the blue filter. Each spectral band which is the emission filter correlated with the red, green, and blue filters has eight classifiers according to the excitation LEDs and the methods. The red dotted circles indicate the classifiers with the highest sensitivity (over 94%).

Figure 13

AI-based band selection method flow chart. Four spectral bands are selected from C01 to C30 in the first stage. The classifier whose AUC is greater than 85% is selected from each selected spectral band. A total of four classifiers are selected in the second stage. The accuracy, sensitivity, and specificity of four selected spectral bands are multiplied by the corresponding weighting score W and summed. The summed accuracy is multiplied by 0.1, the summed sensitivity is multiplied by 0.8, and the summed specificity is multiplied by 0.1. The weighting sore is the summation of the three multiplied values. The method adjusts the weightings to find the maximum weighting score for each selection, and the weighting score of each selection is calculated and compared to find the best filter groups.