Back to search

Aristotle University Of Thessaloniki (AUTH)

Πρόβλεψη της μετάστασης πρωτοπαθούς μελανώματος μέσω αξιολόγησης κλινικών, δερματοσκοπικών και ιστολογικών παραμέτρων, και με τη χρήση νευρωνικού δικτύου τεχνητής νοημοσύνης

Abstract

dc:description

STUDY I Background: Data on predicting melanoma metastasis based on dermatoscopic images of the primary tumor are limited in the literature. Objective: To investigate the association of specific dermatoscopic features of primary melanoma with the risk of developing metastasis. In addition, to develop a dermoscopy-based metastasis prediction model (Model 1) and compare its performance with a model based on Breslow thickness and ulceration (Model 2), as well as a combined model (Model 3). Methods: International, multicenter, retrospective study of melanoma patients (ΑJCC stage IB and above) with available dermatoscopic images of the primary tumor and sufficient follow-up to assess the occurrence of metastasis. Dermatoscopic images were evaluated by experienced raters via an online platform. Models were trained using a stratified split into training and test sets with an 80/20 ratio. The primary outcome was the development of metastasis (locoregional or distant). Secondary outcomes included survival analyses for recurrence-free survival (RFS) and distant metastasis-free survival (DMFS) in patients with early-stage melanoma (I-II). Results: 524 patients were included, 222 (42.4%) developed metastasis at diagnosis or during follow-up (median follow-up 50 months). Dermatoscopic features associated with increased metastatic risk were blue-white veil and extensive dermatoscopic ulceration (>50%). In contrast, heavy pigmentation and extensive regression structures (>50%) showed a protective association. For metastasis prediction, Model 1 predicted metastasis development with an AUC of 0.805, comparable to Model 2, while the combined model had numerically the highest performance. Pairwise comparisons among the three models did not show statistically significant differences in performance. In patients with early-stage tumors, similar dermatoscopic features predicted metastatic risk, and model performance for predicting RFS and DMFS did not differ compared with the models developed in the overall study population. Conclusions: Dermatoscopy may provide prognostic information for melanoma prior to primary tumor excision, with accuracy comparable to the current staging system. Validation of these findings in prospective studies with independent patient populations is required. STUDY II Background: Data on machine learning algorithms employing dermatoscopic images for melanoma prognosis are lacking in the literature. Objective: To investigate the accuracy of a dermatoscopy-based convolutional neural network (CNN) for predicting melanoma metastasis and to compare its performance with a model based on clinicopathological factors (Model 2) and with a combined model (Model 3). Methods: International, multicenter, retrospective study including patients with melanoma stage IA and above, with available dermatoscopic images and sufficient follow-up to determine whether metastasis occurred. A ResNet-50 architecture was used for feature extraction, generating 512-dimensional representations per image. These features were used as input to an SVM classifier (Model 1). A clinicopathological model based on Breslow thickness and histopathologic ulceration (Model 2) and a combined model (SVM + Breslow + ulceration) (Model 3) were also evaluated. Results: A total of 712 patients were included in the study, and 247 developed metastases at diagnosis or during follow-up (median follow-up 60 months). The model based on deep dermatoscopic features achieved strong discrimination (mean AUC 0.84), comparable to Model 2 (mean AUC 0.86), while the combined model showed the highest overall performance (mean AUC 0.87). Differences between models were not statistically significant. In high-risk AJCC groups (e.g., IIB/IIC), Model 1 demonstrated high sensitivity for identifying tumors at increased risk of metastasis, comparable to the staging system. Conclusions: Deep dermatoscopic features can predict metastasis with performance comparable to established clinicopathological prognostic factors. External validation in prospective studies is necessary.

Degree

thesis:*
Grantor dc:publisher
Aristotle University Of Thessaloniki (AUTH)
Year dc:date
2026

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Lallas, Konstantinos
  • Λάλλας, Κωνσταντίνος

Subjects

dc:subject × 24

Rights

Language dc:language
gre

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/61790

Chain of custody

source
Harvested from
Greek National Archive of PhD Theses
Base URL
phdtheses.ekt.gr/eadd_oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
citation

Lallas, Konstantinos; Λάλλας, Κωνσταντίνος. Πρόβλεψη της μετάστασης πρωτοπαθούς μελανώματος μέσω αξιολόγησης κλινικών, δερματοσκοπικών και ιστολογικών παραμέτρων, και με τη χρήση νευρωνικού δικτύου τεχνητής νοημοσύνης. Aristotle University Of Thessaloniki (AUTH), 2026. http://hdl.handle.net/10442/hedi/61790