Histological interpretation of spitzoid tumours: an extensive machine learning-based concordance analysis for...

Area of Science:

• Dermatopathology
• Computational Pathology
• Bioinformatics

Background:

• Histopathological classification of melanocytic tumours with spitzoid features presents diagnostic challenges.
• Interobserver variability can affect the accuracy of classifying these challenging tumours.
• Objective classification criteria are needed to improve diagnostic consistency.

Purpose of the Study:

• To develop and evaluate machine learning (ML) algorithms for objective classification of melanocytic tumours with spitzoid features.
• To identify and rank the most important clinicopathological variables for tumour classification.
• To assess the accuracy and utility of ML algorithms in distinguishing benign, atypical, and malignant spitzoid tumours.

Main Methods:

• Utilized a dataset of 122 melanocytic tumours (benign, atypical, malignant) from four countries.
• Performed analysis of variance (ANOVA) to rank 22 clinicopathological variables.
• Applied Gaussian naive Bayes and logistic regression algorithms for classification tasks.

Main Results:

• Gaussian naive Bayes achieved 0.95 accuracy and 0.87 kappa score distinguishing Spitz naevus from malignant spitzoid tumours.
• For benign versus non-benign Spitz tumours, a kappa score of 0.88 was reached.
• Logistic regression achieved 0.85 accuracy and 0.66 kappa for atypical Spitz tumours versus Spitz melanoma, with most atypical cases misclassified as melanoma.

Conclusions:

• Proposed ML algorithms show promise in supporting histological classification of spitzoid tumours.
• ML can improve accuracy, objectivity, and reduce interobserver variability in tumour diagnosis.
• These algorithms offer a potential solution for the lack of clear classification thresholds for Spitz/spitzoid tumours.