Yuichi Mine1, Shunsuke Suzuki2, Toru Eguchi3, Takeshi Murayama2. 1. Department of Medical System Engineering, Division of Oral Health Sciences, Graduate School of Biomedical and Health Sciences, Hiroshima University, 1-2-3 Kasumi Minami-ku, Hiroshima 734-8553, Japan; Translational Research Center, Hiroshima University, 1-2-3 Kasumi Minami-ku, Hiroshima 734-8553, Japan. Electronic address: mine@hiroshima-u.ac.jp. 2. Department of Medical System Engineering, Division of Oral Health Sciences, Graduate School of Biomedical and Health Sciences, Hiroshima University, 1-2-3 Kasumi Minami-ku, Hiroshima 734-8553, Japan. 3. Graduate School of Engineering, Hiroshima University, 1-3-2 Kagamiyama, Higashi-hiroshima 739-0046, Japan.
Abstract
PURPOSE: Maxillofacial prosthetic rehabilitation replaces missing structures to recover the function and aesthetics relating to facial defects or injuries. Deep learning is rapidly expanding with respect to applications in medical fields. In this study, we apply the artificial neural network (ANN)-based deep learning approach to coloration support for fabricating maxillofacial prostheses. METHODS: We compared two machine learning algorithms, ANN-based deep learning and the random forest algorithm, to determine the compounding amount of pigment. We prepared 52 silicone elastomer specimens of varying colors and measured the CIE 1976 L* a* b* color space information using a spectrophotometer on the input dataset. The output of these algorithms indicated the compounding amount of four pigments. According to the algorithms' pigment compounding predictions, we prepared the specimens for validation analysis and measured the CIE 1976 L* a* b* values. We determined the color differences between the real skin color of five research participants (22.3 ± 1.7 years) and that of the silicone elastomer specimens fabricated based on the algorithm predictions using the CIEDE00 ΔE00 color system. RESULTS: The color differences (ΔE00 value) between the real skin color and silicone elastomer validation specimens were 3.45 ± 0.87 (ANN) and 5.54 ± 1.41 (random forest), which indicates that the deep ANN approach produced superior results with respect to the ΔE00 value compared with the random forest algorithm. CONCLUSIONS: These results suggest that applying deep ANN is a promising technique for the coloration of maxillofacial prostheses.
PURPOSE: Maxillofacial prosthetic rehabilitation replaces missing structures to recover the function and aesthetics relating to facial defects or injuries. Deep learning is rapidly expanding with respect to applications in medical fields. In this study, we apply the artificial neural network (ANN)-based deep learning approach to coloration support for fabricating maxillofacial prostheses. METHODS: We compared two machine learning algorithms, ANN-based deep learning and the random forest algorithm, to determine the compounding amount of pigment. We prepared 52 silicone elastomer specimens of varying colors and measured the CIE 1976 L* a* b* color space information using a spectrophotometer on the input dataset. The output of these algorithms indicated the compounding amount of four pigments. According to the algorithms' pigment compounding predictions, we prepared the specimens for validation analysis and measured the CIE 1976 L* a* b* values. We determined the color differences between the real skin color of five research participants (22.3 ± 1.7 years) and that of the silicone elastomer specimens fabricated based on the algorithm predictions using the CIEDE00 ΔE00 color system. RESULTS: The color differences (ΔE00 value) between the real skin color and silicone elastomer validation specimens were 3.45 ± 0.87 (ANN) and 5.54 ± 1.41 (random forest), which indicates that the deep ANN approach produced superior results with respect to the ΔE00 value compared with the random forest algorithm. CONCLUSIONS: These results suggest that applying deep ANN is a promising technique for the coloration of maxillofacial prostheses.