METHODS: This study utilizes near-infrared hyperspectral imaging, within the range of 900.6 and 1454.8 nm. Images were captured for a total of 125 patients, including 66 patients with Basal Cell Carcinoma, 42 with cutaneous Squamous Cell Carcinoma, and 17 with Actinic Keratosis, to differentiate between healthy and unhealthy skin lesions. A combination of hybrid convolutional neural networks (for feature extraction) and support vector machine algorithms (as a final activation layer) was employed for analysis. In addition, we test whether transfer learning is feasible from networks trained on shorter wavelengths of the electromagnetic spectrum.
RESULTS: The implemented method achieved a general accuracy of over 80%, with some tasks reaching over 90%. F1 scores were also found to generally be over the optimal threshold of 0.8. The best results were obtained when detecting Actinic Keratosis, however differentiation between the two types of malignant lesions was often noted to be more difficult. These results demonstrate the potential of near-infrared hyperspectral imaging combined with advanced machine learning techniques in distinguishing NMSC from other skin lesions. Transfer learning was unsuccessful in improving the training of these algorithms.
CONCLUSIONS: We have shown that the Near-Infrared region of the electromagnetic spectrum is highly useful for the identification and study of non-melanoma type skin lesions. While the results are promising, further research is required to develop more robust algorithms that can minimize the impact of noise in these datasets before clinical application is feasible.
方法:本研究利用近红外高光谱成像,在900.6和1454.8nm的范围内。共捕获125名患者的图像,包括66例基底细胞癌,42患有皮肤鳞状细胞癌,17患有光化性角化病,区分健康和不健康的皮肤病变。采用混合卷积神经网络(用于特征提取)和支持向量机算法(作为最终激活层)的组合进行分析。此外,我们从在电磁波谱的较短波长上训练的网络中测试迁移学习是否可行。
结果:实施的方法达到了80%以上的一般精度,有些任务达到90%以上。还发现F1得分通常超过0.8的最佳阈值。检测光化性角化病时获得了最好的结果,然而,区分这两种类型的恶性病变通常被认为是更困难的。这些结果证明了近红外高光谱成像结合先进的机器学习技术在区分NMSC与其他皮肤病变方面的潜力。迁移学习在改进这些算法的训练方面没有成功。
结论:我们已经表明,电磁波谱的近红外区域对于识别和研究非黑素瘤型皮肤病变非常有用。虽然结果很有希望,需要进一步的研究来开发更稳健的算法,以在临床应用可行之前将这些数据集中的噪声影响降至最低.