In this study, we applied a radiomics approach to skin surface images acquired by a skin analysis device to objectively assess the acute radiation dermatitis in patients undergoing radiotherapy for breast cancer. The 18 radiomic features were calculated from images in the normal, polarised, and UV modes before, during, and after radiotherapy. To test the performance of radiomic features as indicators of radiodermatitis evaluation, we assessed the statistically significant changes and differences within and between groups (ipsilateral and contralateral breasts) over time and correlations between the radiomic features and the RTOG score or skin dose. Furthermore, correlations between the radiomic features at ‘RT D7’ to those at 'RT D14’ and 'after RT D10’ for the ipsilateral breast were analysed to evaluate whether the degree of acute radiation dermatitis can be predicted in advance using the feature.
Figure 2 shows that the skin surface images and the patterns of the corresponding GLCMs in the three imaging modes were changed as the skin was damaged by radiation. Accordingly, most of the radiomic feature ratios for the ipsilateral breasts showed increasing or decreasing tendencies, exhibiting better performance or correlations in all statistical analyses. Of these, the performance of the energy in the normal mode and the sum variance in the polarised and UV modes were better than others in general.
The normal mode in the skin analysis device could analyse skin pores. As shown in Fig. 2(a), it was observed that the skin surface became redder and darker over time on the side of the breast cancer treatment. The results of this study correspond well with those obtained in earlier studies19, 27, 28, 40. Previous studies reported that the skin reaction for irradiated breasts exhibited higher a* values (reddish) and lower L* values (darker); these values were measured by a spectrophotometer during radiotherapy. Momm et al. showed strong correlations to the radiation dose using spectrophotometry. By analysing the spatial distribution of the GLCM intensities in Fig. 2(a), it can be observed that as the skin was damaged by radiation, the peak of the high intensities of GLCMs tended to gradually disappear. These intensities were subsequently shifted toward the lowest elements of the GLCMs. It was demonstrated that intensity homogeneity of the skin surface images deteriorated, and the colour tone of the skin became darker as the radiodermatitis progressed. The energy for the normal imaging mode, which exhibited the best performance for assessing acute radiation dermatitis, had decreasing tendencies over time because it is a measure of homogeneity of an image. A higher value for this feature indicates that the intensity in an image varies less41.
The polarised mode in the skin analysis device could evaluate the melanin pigmentation of the basal layer of the epidermis. Figure 2(b) shows that the skin pigmentation in ipsilateral breasts increased and peaked 'after RT D10.' The results of this study are similar to those of Hu et al.29. Their study reported that the skin pigmentation of irradiated breasts increased, owing to radiation damage, compared with that of unirradiated breasts. The intensities in the GLCM distributions generated from skin surface images in the polarised mode were gradually spread out, and these intensities then shifted toward the lowest elements of the GLCMs, demonstrating that there was inhomogeneity and darkness in the images. The sum variance typically exhibits high values when the intensities of the GLCM distributions are gathered in the lowest and highest elements38. For this reason, the values of the sum variance decreased over time in this study.
The UV mode in the skin analysis device was typically utilised to evaluate skin sebum; however, it also showed melanin accumulations below the skin surface. In this study, the skin surface images in the UV mode continued to darken from the periphery inward, toward the centre of the images, over time. This demonstrated that the melanin accumulated, owing to radiation damage, as shown in Fig. 2(c). It was difficult to evaluate radiodermatitis using skin sebum. Hu et al. also showed that there was no significant change in skin sebum in both ipsilateral and contralateral breasts before and after radiotherapy29. Unlike other imaging modes, the periphery of the images in the UV mode appeared optically darker than the centre, owing to non-uniformity in UV intensity. For this reason, we cropped the circular ROI of the original images obtained in UV mode. Other images in the normal and polarised mode were also cropped in the same manner to calculate the radiomic features. The intensities in the GLCM distributions generated from skin surface images in the UV mode exhibited a pattern similar to those in the polarised mode. Therefore, the sum variance also had decreasing tendencies, providing the best performance for both polarised and UV imaging modes.
Figure 3 presents the changes in the ratios of the representative radiomic features for ipsilateral and contralateral breasts over time in the three imaging modes. As mentioned above, these radiomic features showed the statistical significance of the two-way repeated measures ANOVA test. In the unirradiated breasts, there were unpredictable tendencies in the energy in the normal mode. There were decreasing tendencies in the sum variance in both the polarised and UV modes; however, these showed a much smaller decrease compared to the irradiated breasts. This may be explained by the fact that radiation that was used to treat breast cancer was scattered or penetrated to the contralateral breast, thereby causing damage. Hu et al. reported similar results, showing an increase in pigmentation for the contralateral breast after radiotherapy29. According to the results of our study, radiation-induced skin damage for contralateral breasts ‘after RT D10’ was comparable to that for ipsilateral breasts on ‘RT D7’. Thus, the degree of radiation-induced skin damage to the unirradiated breast could be quantified and evaluated using radiomics analysis of the skin surface image proposed in this study.
Several studies have addressed the limitation of these subjective methods, which create several uncertainties. They demonstrated that these methods could not represent patient-reported breast symptoms, including pain, itching, tightness, and local heat42. The subjective assessments were affected by intra- and inter-evaluator variations, and were considered less suitable for close scrutiny15, 22, 31. Although these various studies highlighted the limitations of the clinician-assessed scoring criteria, the performances of various skin biophysical parameters as indicators for assessing radiodermatitis were evaluated by analysing the correlations between the scoring criteria and the parameters. In this study, we used the RTOG scoring criteria, and all enrolled patients were evaluated as grade 1 on ‘RT D14’ and ‘after RT D10.’ For this reason, our study could not show the correlations between the RTOG scoring and the radiomic feature ratios for different time points. However, we presented strong correlations between the skin doses and radiomic feature ratios on ‘RT D7.’ Thus, the severity of acute radiation dermatitis can vary within the same grade, showing the limitations of clinician-assessed scoring criteria. Because there is a significant relationship between the severity of radiodermatitis and the absolute doses delivered to the patients’ skin6, 7, we could evaluate the severity of acute radiation dermatitis objectively using representative radiomic features (on 'RT D7'), regardless of the clinical-assessed scoring criteria. Furthermore, the representative radiomic features exhibited strong correlations to those on 'RT D14' and 'after RT D10,’ which means that radiation-induced skin damage during and after radiotherapy can be predicted in advance. Using a radiomics approach to skin surface images, it is possible to subdivide the severity of acute radiation dermatitis even within the same grade and then prepare appropriate countermeasures.
In future work, this radiomic approach for the evaluation of radiodermatitis will be further applied to patients undergoing radiotherapy for head and neck cancer, where acute radiation dermatitis is predominantly observed. Because patients over grade 2+ were not included, we were unable to evaluate the comprehensive severity of radiodermatitis, which is a limitation of this study. Further research using a large number of samples and various severities of radiodermatitis, various treatment sites, and various races will be performed in the future. In addition to the hand-crafted radiomic features utilised in our study, quantitative radiomic features extracted using a convolutional neural network can be used in supervised machine learning to predict the radiodermatitis25.
A radiomics approach to skin surface images acquired by a skin analysis device was applied to objectively assess acute radiation dermatitis in patients undergoing radiotherapy for breast cancer. In general, the energy for the normal mode and sum variance for polarised and UV modes exhibited better performance than others in evaluating the radiation-induced skin damage. Using the ratios of the energy and sum variance on 'RT D7,’ radiodermatitis severity during and after radiotherapy can be predicted in advance, which assists in its appropriate management.