A deep learning-clinical nomogram hybrid for predicting sentinel lymph node metastasis in melanoma
Minh-Khang Le1, Masataka Kawai1, Tetsuo Kondo1
1Department of Pathology, University of Yamanashi, Yamanashi, Japan.
Background:
Melanoma is a life-threatening skin malignancy, with sentinel lymph node metastasis (SLNM) serving as a critical prognostic factor. While machine and deep learning models using histopathology have focused on melanoma diagnosis, limited efforts have aimed to predict SLNM.
Objectives:
This study aimed to develop a collaborative machine and deep learning model that integrates histopathological patches and clinical data to predict SLNM in patients with invasive melanoma, thereby aiding decision-making regarding sentinel lymph node biopsy.
Methods:
A total of 78 invasive cutaneous melanoma cases (43 SLNM-positive and 35 SLNM-negative) from the University of Yamanashi Hospital (2007-2022) were divided into training (n = 60) and test (n = 18) datasets. A clustering-constrained attention multiple-instance learning (CLAM) model was developed using haematoxylin and eosin-stained whole-slide images. Among the four CLAM variants, the model using a ResNet50trunc encoder trained on whole-tissue regions (CLAM-R50) demonstrated the best performance and was further integrated with a clinical nomogram. Hybrid modelling was performed using logistic regression, a support vector machine and gradient boosting.
Results:
CLAM-R50 achieved an area under the receiver operating characteristic curve (AUROC) of 0.875, while the clinical nomogram achieved an AUROC of 0.826 on the test set. The gradient boosting hybrid model, termed the Melanoma Indicative Scorer for Sentinel Lymph node Evaluation (MISSLE), demonstrated the highest AUROC of 0.950 (95% bootstrapped confidence interval (CI): 0.933-1.00) on the test set. CLAM-R50 attention maps identified pale cytoplasm and plasma cell-like tumour cells as predictors of positive SLNM, whereas dense melanin deposition and lymphocyte-like tumour cells predicted negative SLNM.
Conclusions:
MISSLE integrates histopathological and clinical data to predict SLNM with high AUROC, surpassing the previous reports and offering a tool for clinical decision-making. Limitations include the single-centre design and predominantly Japanese cohort, limiting generalizability. The code and models are publicly available for external validation.
