Artificial Neural Networks (ANN) for automatic detection of dendritic-shaped cancer cells of cutaneous melanoma in Reflectance Confocal Microscopy (RCM) images

Melanoma (MM) is one of the tumors with the highest incidence. In Italy, MM affected about 13,700 patients out of 373,000 new cases of cancer in 2018, with prognosis dependent on the degree of tumor invasion and presence of metastasis at diagnosis: only an early detection can lead to a better prognosis. Recent evidence suggests that MM is a family of different tumors with varying abilities to grow and metastasize: dendritic-shaped tumor cells were typically found in thin MM in situ. Reflectance Confocal Microscopy (RCM) is a non-invasive imaging tool that enables in vivo observation of the skin at a quasi-histological resolution, providing transverse-section grayscale images related to refractive index of different tissues. In this work, a dataset of RCM images, from 13 healthy subjects and 22 patients affected by MM in situ, were used to train a Multi-Layer Perceptron (MLP) artificial neural network. Each image was subdivided into sub-blocks, labeled as positive if containing significant clusters of dendritic-shaped tumour cells. In each block, various standard features were calculated, e.g. Haralick's and features from the run-length matrices. The MLP was trained to recognize the presence of clusters of dendritic-shaped cancer cells. The preliminary results are encouraging, giving AUC=0.81 with about 73% accuracy. Tests are currently underway to improve quality.

黑色素瘤（Melanoma，简称MM）是发病率最高的肿瘤之一。2018年，意大利37.3万例新发癌症病例中，约有1.37万例为MM患者；其预后取决于诊断时肿瘤的侵袭程度及是否存在转移，唯有早期检测方能改善预后。近期研究表明，MM是一类具有不同生长和转移能力的异质性肿瘤家族；树突状肿瘤细胞通常见于表浅的原位MM中。反射式共聚焦显微镜（Reflectance Confocal Microscopy，简称RCM）是一种无创成像工具，可在近似组织学分辨率下对皮肤进行在体观察，并提供与不同组织折射率相关的横断面灰度图像。本研究使用了来自13名健康受试者和22名原位MM患者的RCM图像数据集，以训练多层感知器（Multi-Layer Perceptron，简称MLP）人工神经网络。每张图像被划分为多个子块，若子块中包含显著的树突状肿瘤细胞簇，则标记为阳性。在每个子块中计算了多种标准特征，例如Haralick特征和游程矩阵特征。该MLP模型经训练后可识别树突状癌细胞簇的存在。初步结果令人鼓舞，其AUC值为0.81，准确率约为73%。目前正在进行测试以提升模型性能。