Department of Dermatology and Dermatologic Surgery, Medical University of South Carolina, Charleston, SC.
Clemson University School of Computing, Clemson, SC.
Am J Dermatopathol. 2022 Sep 1;44(9):650-657. doi: 10.1097/DAD.0000000000002232. Epub 2022 Jul 19.
The integration of an artificial intelligence tool into pathologists' workflow may lead to a more accurate and timely diagnosis of melanocytic lesions, directly patient care. The objective of this study was to create and evaluate the performance of such a model in achieving clinical-grade diagnoses of Spitz nevi, dermal and junctional melanocytic nevi, and melanomas.
We created a beginner-level training environment by teaching our algorithm to perform cytologic inferences on 136,216 manually annotated tiles of hematoxylin and eosin-stained slides consisting of unequivocal melanocytic nevi, Spitz nevi, and invasive melanoma cases. We sequentially trained and tested our network to provide a final diagnosis-classification on 39 cases in total. Positive predictive value (precision) and sensitivity (recall) were used to measure our performance.
The tile-classification algorithm predicted the 136,216 irrelevant, melanoma, melanocytic nevi, and Spitz nevi tiles at sensitivities of 96%, 93%, 94% and 73%, respectively. The final trained model was able to correctly classify and predict the correct diagnosis in 85.7% of unseen cases (n = 28), reporting at or near screening-level performances for precision and recall of melanoma (76.2%, 100.0%), melanocytic nevi (100.0%, 75.0%), and Spitz nevi (100.0%, 75.0%).
Our pilot study proves that convolutional networks trained on cellular morphology to classify melanocytic proliferations can be used as a powerful tool to assist pathologists in screening for melanoma versus other benign lesions.
将人工智能工具集成到病理学家的工作流程中，可以更准确、及时地诊断黑色素细胞病变，直接改善患者的护理效果。本研究的目的是创建和评估此类模型在实现 Spitz 痣、真皮和交界性黑色素细胞痣以及黑色素瘤的临床诊断级别的性能。
我们通过教授算法对 136216 张经人工注释的苏木精和伊红染色载玻片进行细胞学推断，创建了一个初级培训环境，这些载玻片包含明确的黑色素细胞痣、Spitz 痣和侵袭性黑色素瘤病例。我们依次对我们的网络进行训练和测试，总共对 39 个病例提供最终的诊断分类。使用阳性预测值（精度）和敏感性（召回率）来衡量我们的性能。
瓦片分类算法对 136216 张无关的、黑色素瘤、黑色素细胞痣和 Spitz 痣的瓦片的敏感性分别为 96%、93%、94%和 73%。最终训练好的模型能够正确分类并预测 28 个未见过的病例中的 85.7%（n=28），在黑色素瘤（76.2%，100.0%）、黑色素细胞痣（100.0%，75.0%）和 Spitz 痣（100.0%，75.0%）的准确性和召回率方面，报告的结果接近筛查水平。
我们的初步研究证明，经过细胞形态学训练的卷积网络可以用于分类黑色素细胞增生，作为一种强大的工具，帮助病理学家筛查黑色素瘤与其他良性病变。
