Introduction: the incidence of melanoma is continuously increasing and it is responsible for the majority of skin cancer deaths. Early diagnosis and complete removal of the tumor tissue before the onset of deep invasion are the main factors for the reduction of its mortality and morbidity. For this reason, different non-invasive imaging techniques to allow the early and more accurate identification of malignant lesions have been developed. Objectives of the study: this study aims to elaborate an AI integrated approach, based on images made using Line-field confocal optical coherence tomography (LC-OCT), able to correctly identify dermoscopically suspicious melanocytic lesions as benign, malignant or at uncertain prognostic significance. Materials and methods: A retrospective study was conducted to elaborate and evaluate the accuracy of an AI model in distinguishing images of melanocytic lesions into three diagnostic categories: melanocytic nevi, atypical/dysplastic melanocytic nevi and melanoma. The analyzed images consisted of vertical sections (DICOM) acquired via LC-OCT. The set of LC-OCT DICOM images was enriched, in order to try to increase the signal/noise ratio (SNR) with respect to the investigation outcome with different filters (RAW, Gaussian, LOG and MERGED). The images filtered with Gaussian and RAW filter were used to extract pixel clusters, according to the SLIC Superpixels and Affinity Propagation (AP) Clustering algorithm. Regions of interest were extracted and image biomarkers were extracted with a library in R, called Moddicom. Different machine learning models (bivariate logistic regression, decision tree and random forest) were trained to understand if any biomarker was able to discern melanoma from benign moles. Results: 127 variables showed a statistically significant p-value on univariate testing. The most promising bivariate regression models were extracted. The performances on the training set are high, while those on the testing set are lower: this means that apparently good results of this technique are probably due to overfitting. Similarly, the decision tree and the random forest also showed excellent levels of accuracy, positive predictive value (PPV) and negative predictive value (NPV) for the training test with a drop in performance on the independent internal testing set. Discussion and conclusions: In this pilot study we developed and demonstrated, for the first time in the literature, the feasibility of an artificial intelligence model for the in-vivo discrimination between benign melanocytic lesions, uncertain malignant potential lesions and melanoma based on LC-OCT images. We have identified several biomarkers potentially useful for this purpose. Although this model has not yet demonstrated good results in terms of performance, several ways for its improvement have been identified and proposed.
Introduzione: l'incidenza del melanoma è in continuo aumento ed è responsabile della maggior parte dei decessi per tumore della pelle. La diagnosi precoce e la rimozione completa del tumore prima che inizi la fase di invasione profonda costituiscono le principali armi per ridurne mortalità e morbidità. Per questo motivo sono state sviluppate diverse tecniche non invasive di imaging per consentire l'identificazione precoce e più accurata delle lesioni maligne. Obiettivi dello studio: questo studio si propone di elaborare un modello basato sull’ utilizzo dell’intelligenza artificiale (AI), in grado di distinguere, partendo da immagini ottenute con la tomografia a coerenza ottica confocale (LC-OCT), lesioni melanocitiche sospette all’indagine dermoscopica in benigne, maligne o con significato prognostico incerto. Materiali e metodi: è stato condotto uno studio retrospettivo per elaborare e valutare l'accuratezza di un modello AI nel distinguere le immagini di lesioni melanocitiche in tre categorie diagnostiche: nevi melanocitici, nevi melanocitici atipici/displastici e melanoma. Sono state analizzate sezioni verticali (DICOM standard) acquisite tramite LC-OCT. Il set di immagini LC-OCT è stato arricchito, per cercare di aumentare il rapporto segnale/rumore (SNR) con diversi filtri (RAW, Gaussiano, LOG e MERGED). Le immagini filtrate con filtro gaussiano e RAW sono state utilizzate per estrarre cluster di pixel, secondo l'algoritmo SLIC Superpixels e Affinity Propagation (AP) Clustering. Sono state estratte le regioni di interesse e i biomarcatori con una libreria in R, chiamata Moddicom. Diversi modelli di machine learning (regressione logistica bivariata, albero decisionale e random forest) sono stati addestrati per capire se qualche biomarcatore fosse in grado di distinguere il melanoma dalle altre lesioni. Discussione e conclusioni: In questo studio pilota abbiamo sviluppato e dimostrato, per la prima volta in letteratura, la possibilità di sviluppare un modello di intelligenza artificiale per la discriminazione in vivo tra lesioni melanocitiche benigne, lesioni a potenziale maligno incerto e melanoma, basato su immagini LC-OCT. Abbiamo identificato diversi biomarcatori potenzialmente utili a questo scopo. Sebbene questo modello non abbia ancora dimostrato buoni risultati in termini di prestazioni, sono stati identificati e proposti diverse strade per migliorarlo.
ARTIFICIAL INTELLIGENCE-BASED INTEGRATED APPROACH FOR SKIN CANCER RECOGNITION
SOGLIA, SIMONE
2025
Abstract
Introduction: the incidence of melanoma is continuously increasing and it is responsible for the majority of skin cancer deaths. Early diagnosis and complete removal of the tumor tissue before the onset of deep invasion are the main factors for the reduction of its mortality and morbidity. For this reason, different non-invasive imaging techniques to allow the early and more accurate identification of malignant lesions have been developed. Objectives of the study: this study aims to elaborate an AI integrated approach, based on images made using Line-field confocal optical coherence tomography (LC-OCT), able to correctly identify dermoscopically suspicious melanocytic lesions as benign, malignant or at uncertain prognostic significance. Materials and methods: A retrospective study was conducted to elaborate and evaluate the accuracy of an AI model in distinguishing images of melanocytic lesions into three diagnostic categories: melanocytic nevi, atypical/dysplastic melanocytic nevi and melanoma. The analyzed images consisted of vertical sections (DICOM) acquired via LC-OCT. The set of LC-OCT DICOM images was enriched, in order to try to increase the signal/noise ratio (SNR) with respect to the investigation outcome with different filters (RAW, Gaussian, LOG and MERGED). The images filtered with Gaussian and RAW filter were used to extract pixel clusters, according to the SLIC Superpixels and Affinity Propagation (AP) Clustering algorithm. Regions of interest were extracted and image biomarkers were extracted with a library in R, called Moddicom. Different machine learning models (bivariate logistic regression, decision tree and random forest) were trained to understand if any biomarker was able to discern melanoma from benign moles. Results: 127 variables showed a statistically significant p-value on univariate testing. The most promising bivariate regression models were extracted. The performances on the training set are high, while those on the testing set are lower: this means that apparently good results of this technique are probably due to overfitting. Similarly, the decision tree and the random forest also showed excellent levels of accuracy, positive predictive value (PPV) and negative predictive value (NPV) for the training test with a drop in performance on the independent internal testing set. Discussion and conclusions: In this pilot study we developed and demonstrated, for the first time in the literature, the feasibility of an artificial intelligence model for the in-vivo discrimination between benign melanocytic lesions, uncertain malignant potential lesions and melanoma based on LC-OCT images. We have identified several biomarkers potentially useful for this purpose. Although this model has not yet demonstrated good results in terms of performance, several ways for its improvement have been identified and proposed.File | Dimensione | Formato | |
---|---|---|---|
Tesi Dottorato.pdf
accesso aperto
Dimensione
7.15 MB
Formato
Adobe PDF
|
7.15 MB | Adobe PDF | Visualizza/Apri |
I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/20.500.14242/202555
URN:NBN:IT:UNIBS-202555