Laboratorio di Biostatistica e Bioinformatica, I.R.C.C.S. Istituto Tumori 'Giovanni Paolo II', Bari, Italy.
Unità Operativa Complessa di Anatomia Patologica, I.R.C.C.S. Istituto Tumori "Giovanni Paolo II", Bari, Italy.
J Transl Med. 2024 Sep 12;22(1):838. doi: 10.1186/s12967-024-05629-2.
Risk stratification and treatment benefit prediction models are urgent to improve negative sentinel lymph node (SLN-) melanoma patient selection, thus avoiding costly and toxic treatments in patients at low risk of recurrence. To this end, the application of artificial intelligence (AI) could help clinicians to better calculate the recurrence risk and choose whether to perform adjuvant therapy.
We made use of AI to predict recurrence-free status (RFS) within 2-years from diagnosis in 94 SLN- melanoma patients. In detail, we detected quantitative imaging information from H&E slides of a cohort of 71 SLN- melanoma patients, who registered at Istituto Tumori "Giovanni Paolo II" in Bari, Italy (investigational cohort, IC). For each slide, two expert pathologists firstly annotated two Regions of Interest (ROIs) containing tumor cells alone (TUMOR ROI) or with infiltrating cells (TUMOR + INF ROI). In correspondence of the two kinds of ROIs, two AI-based models were developed to extract information directly from the tiles in which each ROI was automatically divided. This information was then used to predict RFS. Performances of the models were computed according to a 5-fold cross validation scheme. We further validated the prediction power of the two models on an independent external validation cohort of 23 SLN- melanoma patients (validation cohort, VC).
The TUMOR ROIs have revealed more informative than the TUMOR + INF ROIs. An Area Under the Curve (AUC) value of 79.1% and 62.3%, a sensitivity value of 81.2% and 76.9%, a specificity value of 70.0% and 43.3%, an accuracy value of 73.2% and 53.4%, were achieved on the TUMOR and TUMOR + INF ROIs extracted for the IC cohort, respectively. An AUC value of 76.5% and 65.2%, a sensitivity value of 66.7% and 41.6%, a specificity value of 70.0% and 55.9%, an accuracy value of 70.0% and 56.5%, were achieved on the TUMOR and TUMOR + INF ROIs extracted for the VC cohort, respectively.
Our approach represents a first effort to develop a non-invasive prognostic method to better define the recurrence risk and improve the management of SLN- melanoma patients.
风险分层和治疗获益预测模型对于改善阴性前哨淋巴结（SLN-）黑色素瘤患者的选择至关重要，从而避免低复发风险患者接受昂贵且有毒的治疗。为此，人工智能（AI）的应用可以帮助临床医生更好地计算复发风险，并选择是否进行辅助治疗。
我们利用 AI 预测 94 例 SLN-黑色素瘤患者诊断后 2 年内无复发生存（RFS）状态。具体来说，我们从意大利巴里 Istituto Tumori“Giovanni Paolo II”登记的 71 例 SLN-黑色素瘤患者的 HE 切片中检测定量成像信息（研究队列，IC）。对于每张切片，两位专家病理学家首先注释了包含肿瘤细胞的两个感兴趣区（ROI）（肿瘤 ROI）或包含浸润细胞的 ROI（肿瘤+浸润 ROI）。对于两种 ROI，开发了两种基于 AI 的模型，从自动划分每个 ROI 的瓦片中提取信息。然后，使用该信息来预测 RFS。根据 5 折交叉验证方案计算模型的性能。我们还在 23 例 SLN-黑色素瘤患者的独立外部验证队列（验证队列，VC）上验证了两种模型的预测能力。
肿瘤 ROI 比肿瘤+浸润 ROI 提供了更有价值的信息。在 IC 队列中，从肿瘤和肿瘤+浸润 ROI 中提取的 AUC 值分别为 79.1%和 62.3%，灵敏度值分别为 81.2%和 76.9%，特异性值分别为 70.0%和 43.3%，准确率分别为 73.2%和 53.4%。在 VC 队列中，从肿瘤和肿瘤+浸润 ROI 中提取的 AUC 值分别为 76.5%和 65.2%，灵敏度值分别为 66.7%和 41.6%，特异性值分别为 70.0%和 55.9%，准确率分别为 70.0%和 56.5%。
我们的方法代表了开发一种非侵入性预后方法的首次尝试，以更好地定义复发风险并改善 SLN-黑色素瘤患者的管理。
