Background: Breast canceris the most common malignancy among women and one of the leading causes of cancer-related death worldwide. Early diagnosis through mammography remains essential to improve prognosis and survival rates. However, the increasing complexity of imaging data and the variability in human interpretation highlight the need for advanced quantitative tools to support clinical decision-making. In this context, radiomics and deep learning have emerged as promising methodologies, enabling the extraction of quantitative features from medical images that can capture subtle tissue heterogeneity invisible to the human eye. Aim of the study: The aim of this study was to compare the performance of a radiomic workflow based on Machine Learning (ML)algorithms with that of a Deep Learning (DL) model for the automatic classification of breast lesions, assessing their potential as diagnostic support tools in breast imaging. Materials and methods: For internal validation, the CBIS-DDSM (Curated Breast Imaging Subset of the Digital Database for Screening Mammography) dataset was used, publicly available through the Cancer Imaging Archive (TCIA). The database includes digitized mammographic images annotated with segmentation masks and histopathological reports, divided into two groups: microcalcifications (1,872 images corresponding to 753 patients) and masses (1,696 images corresponding to 891 patients). Among these 3568 images, 2111 cases are benign, while 1457 are malignant. This dataset was integrated with an external dataset from the “G. Giglio” Institute of Cefalù (Italy), consisting of 222 independent mammographic images from 180 female patients, all enrolled in a real clinical setting between February 2021 and December 2024. This external dataset enabled the assessment of the models’ generalizability beyond the initial study environment. In the radiomic workflow, images were preprocessed (contras tequalization and noise reduction), and several morphological and texture features were extracted from each lesion. The most relevant variables were selected to train classification models based on Linear Discriminant Analysis (LDA) and Support Vector Machine (SVM). The Deep Learning model (EfficientNetB6), pre-trained on ImageNet, was retrained through transfer learning and fine-tuning strategies. Additionally, data augmentation techniques were applied to reduce overfitting. Model performance was evaluated using AUC, accuracy, sensitivity, specificity, and F1-score. Results and conclusion: The ML-based models achieved an AUC of 68.3% for microcalcifications and 61.5% for masses, while the DL model reached81.5% for masses and 76.2% for microcalcifications, demonstrating superior discriminative power and better external generalization. These findings confirm the potential of Deep Learning as an effective decision-support tool for radiologists, capable of improving diagnostic accuracy and consistency. Nonetheless, traditional radiomics remains complementary, and the integration of hybrid AI models may represent the most promising direction to ward quantitative and precision radiology.
Introduzione: Il carcinoma mammario rappresenta la neoplasia più frequente nella popolazione femminile e una delle principali cause di mortalità oncologica. La diagnosi precoce, basata sull’identificazione radiologica di lesioni caratteristiche come masse e microcalcificazioni all’esame mammografico è fondamentale per migliorare la prognosi e ridurre la mortalità. Tuttavia, l’elevata complessità delle immagini e la variabilità interpretativa rendono necessario l’impiego di strumenti quantitativi di supporto alla diagnosi. In tale contesto, la radiomica e il deep learning si sono affermati come approcci promettenti per l’analisi avanzata delle immagini mediche, consentendo di estrarre ed interpretare caratteristiche quantitative (features) altrimenti non visibili all’occhio umano. Obiettivo dello studio: Scopo del presente lavoro è stato confrontare le prestazioni di worfkflowradiomico basato su algoritmi di Machine Learning (ML) con quelle di (DL) nella classificazione automatica delle lesioni mammarie, al fine di valutare il contributo delle metodiche di intelligenza artificiale nella diagnosi radiologica del carcinoma mammario. Materiali e metodi: Per la validazione interna è stato utilizzato il dataset CBIS-DDSM (CuratedBreast Imaging Subset of Digital Database for Screening Mammography), disponibile pubblicamente tramite il Cancer imaging archive (TCIA). Il database comprende immagini mammografiche digitalizzate annotate con maschere di segmentazione e referti istologici,suddivise in due gruppi: microcalcificazioni (1872 immagini corrispondenti a 753 pazienti) e masse (1696 immagini corrispondenti a 891 pazienti). Tra queste 3568 immagini, 2111 casi sono benigni, mentre 1457 sono maligni. IL suddetto dataset è stato integrato con un dataset esterno dell’Istituto “G.Giglio” di Cefalù, costituito da 222 immagini mammografiche indipendenti, relative a 180 pazienti, tutte donne, arruolate in un contesto clinico reale, da febbraio 2021 a dicembre 2024. Questo dataset esterno ha permesso di testare la generalizzabilità dei modelli al di fuori del contesto di studio iniziale. Nel workflow radiomico, le immagini sono state preprocessate (equalizzazione del contrasto e riduzione del rumore) e da ciascuna lesione sono state estratte numerose features morfologiche e di texture. Le variabili più rilevanti sono state selezionate per addestrare modelli di classificazione basati su Linear Discriminant Analysis (LDA) e Support Vector Machine (SVM). Il modello Deep Learning (EfficientNetB6), preaddestrato su Image-Net, è stato riaddestrato mediante strategie di trasferlearnng e fine-tuning; sono state, inoltre, applicate tecniche di data augmentation per ridurre l’overfitting. Le performance sono state valutate tramite AUC, accuratezza, sensibilità, specificità e F1-score. Risultati e conclusioni: I modelli ML hanno raggiunto un’AUC del 68,3% per le microcalcificazioni e del 61,5% per le masse. Il modello DL ha ottenuto, invece, un’AUC dell’81% per le masse e del 76,2% per le calcificazioni, mostrando una superiore capacità discriminativa e una migliore generalizzazione in validazione esterna. Questi risultati confermano il potenziale del Deep Learning come strumento di supporto al radiologo, in grado di incrementare la sensibilità diagnostica e ridurre la variabilità interosservatore. Tuttavia, la radiomica tradizionale mantiene un ruolo complementare e l’integrazione di approcci ibridi rappresenta la prospettiva più promettente verso una radiologia quantitativa e personalizzata.
VALUTAZIONE COMPARATIVA TRA RADIOMICA BASATA SU MACHINE LEARNING E DEEP LEARNING PER LA CLASSIFICAZIONE DELLE LESIONI MAMMARIE IN MAMMOGRAFIA
NARBONESE, DANIELA
2023/2024
Abstract
Background: Breast canceris the most common malignancy among women and one of the leading causes of cancer-related death worldwide. Early diagnosis through mammography remains essential to improve prognosis and survival rates. However, the increasing complexity of imaging data and the variability in human interpretation highlight the need for advanced quantitative tools to support clinical decision-making. In this context, radiomics and deep learning have emerged as promising methodologies, enabling the extraction of quantitative features from medical images that can capture subtle tissue heterogeneity invisible to the human eye. Aim of the study: The aim of this study was to compare the performance of a radiomic workflow based on Machine Learning (ML)algorithms with that of a Deep Learning (DL) model for the automatic classification of breast lesions, assessing their potential as diagnostic support tools in breast imaging. Materials and methods: For internal validation, the CBIS-DDSM (Curated Breast Imaging Subset of the Digital Database for Screening Mammography) dataset was used, publicly available through the Cancer Imaging Archive (TCIA). The database includes digitized mammographic images annotated with segmentation masks and histopathological reports, divided into two groups: microcalcifications (1,872 images corresponding to 753 patients) and masses (1,696 images corresponding to 891 patients). Among these 3568 images, 2111 cases are benign, while 1457 are malignant. This dataset was integrated with an external dataset from the “G. Giglio” Institute of Cefalù (Italy), consisting of 222 independent mammographic images from 180 female patients, all enrolled in a real clinical setting between February 2021 and December 2024. This external dataset enabled the assessment of the models’ generalizability beyond the initial study environment. In the radiomic workflow, images were preprocessed (contras tequalization and noise reduction), and several morphological and texture features were extracted from each lesion. The most relevant variables were selected to train classification models based on Linear Discriminant Analysis (LDA) and Support Vector Machine (SVM). The Deep Learning model (EfficientNetB6), pre-trained on ImageNet, was retrained through transfer learning and fine-tuning strategies. Additionally, data augmentation techniques were applied to reduce overfitting. Model performance was evaluated using AUC, accuracy, sensitivity, specificity, and F1-score. Results and conclusion: The ML-based models achieved an AUC of 68.3% for microcalcifications and 61.5% for masses, while the DL model reached81.5% for masses and 76.2% for microcalcifications, demonstrating superior discriminative power and better external generalization. These findings confirm the potential of Deep Learning as an effective decision-support tool for radiologists, capable of improving diagnostic accuracy and consistency. Nonetheless, traditional radiomics remains complementary, and the integration of hybrid AI models may represent the most promising direction to ward quantitative and precision radiology.| File | Dimensione | Formato | |
|---|---|---|---|
|
Narbonese_Daniela.pdf
Accesso riservato
Dimensione
869.23 kB
Formato
Adobe PDF
|
869.23 kB | Adobe PDF |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/97587