A deep learning-clinical nomogram hybrid for predicting sentinel lymph node metastasis in melanoma

Area of Science:

• Oncology
• Computational Pathology
• Medical Informatics

Background:

• Melanoma is a dangerous skin cancer where metastasis to lymph nodes is a key prognostic indicator.
• Predicting sentinel lymph node metastasis (SLNM) in melanoma is crucial for patient management but has been under-explored using computational methods.
• Existing machine and deep learning models primarily focus on melanoma diagnosis, not SLNM prediction.

Purpose of the Study:

• To develop a collaborative machine and deep learning model integrating histopathology and clinical data for predicting SLNM in invasive melanoma.
• To enhance clinical decision-making regarding sentinel lymph node biopsy (SLNB).

Main Methods:

• Developed a clustering-constrained attention multiple-instance learning (CLAM) model using H&E-stained whole-slide images from 78 invasive melanoma cases.
• Utilized a ResNet50 encoder (CLAM-R50) and integrated it with a clinical nomogram.
• Created a hybrid model using logistic regression, SVM, and gradient boosting, named Melanoma Indicative Scorer for Sentinel Lymph node Evaluation (MISSLE).

Main Results:

• CLAM-R50 achieved an AUROC of 0.875; the clinical nomogram achieved 0.826.
• The MISSLE hybrid model demonstrated superior performance with an AUROC of 0.950 on the test set.
• Attention maps identified specific cellular features associated with SLNM status.

Conclusions:

• MISSLE effectively integrates histopathological and clinical data to predict SLNM with high accuracy, outperforming previous methods.
• This model serves as a valuable tool for clinical decision-making in melanoma management.
• Future validation in diverse cohorts is needed due to single-center and cohort limitations.