https://hdl.handle.net/10807/60 2026-06-02T06:38:04Z 2026-06-02T06:38:04Z https://hdl.handle.net/10807/337465 2026-06-02T00:09:29Z 2025-01-01T00:00:00Z

Titolo: Artificial Intelligence Techniques for Dental Artifact Suppression in Medical Imaging Autori: Frasca, Maria; Lin, Jianyi; La Torre, Davide Abstract: Dental artifacts are among the most significant diagnostic challenges in computed tomography (CT) and magnetic resonance imaging (MRI), as they degrade image quality and compromise clinical interpretation. This systematic review examined 15 original studies, selected according to the PRISMA flow diagram, with the objective of identifying and evaluating artificial intelligence (AI) algorithms developed for the detection and mitigation of dental artifacts. Performance was assessed with respect to both image quality enhancement, measured through quantitative metrics such as PSNR and SSIM, and diagnostic accuracy. The methodological rigor of the included studies was appraised using the Newcastle–Ottawa Scale. The findings reveal an increasing adoption of deep learning methods—including convolutional neural networks (CNNs), U-Net architectures, and Transformer-based models—alongside traditional metal artifact reduction (MAR) techniques. While these approaches show encouraging results, their performance remains heterogeneous and constrained by the limited size of available datasets. Based on the evidence, we propose a preliminary methodological pipeline to integrate advanced AI techniques for artifact removal and subsequent image classification, with the aim of improving both image quality and diagnostic reliability.

2025-01-01T00:00:00Z https://hdl.handle.net/10807/337464 2026-06-02T00:10:17Z 2025-01-01T00:00:00Z

Titolo: Self-supervised learning model leveraging structural similarity index for glaucoma recognition Autori: Frasca, Maria; Lin, Jianyi; La Torre, Davide Abstract: Glaucoma is a leading cause of irreversible blindness worldwide, affecting millions annually. Early diagnosis is crucial but challenging due to the asymptomatic nature of the disease and the complexity of clinical signs. Advanced ocular imaging generates complex, voluminous data, demanding significant expertise for interpretation. Artificial intelligence (AI) is transforming this field by enabling efficient data analysis. We present a comprehensive approach that combines advanced image preprocessing techniques, leveraging the Discrete Fourier Transform (DFT) implemented via the Fast Fourier Transform (FFT) algorithm, with state-of-the-art AI models to enhance diagnostic accuracy. Using the REFUGE2 dataset, this pre-processing achieved high visual fidelity, with SSIM values of 0.9225 for training and 0.8247 for testing. For classification, we employed Transformer architectures, including Masked Image Modeling (MIM) and Masked Autoencoder (MAE), integrated with Self-Supervised Learning (SimCLR). This multi-model approach yielded an accuracy of 90.06% with MIM and 90.14% with MAE on preprocessed images, surpassing the performance achieved on non-preprocessed data. The MIM model demonstrated a strong discriminative capability, highlighted by an AUC-ROC of 0.9298. Our findings show that combining robust preprocessing with Transformer networks significantly improves efficiency and diagnostic accuracy in the early detection of glaucoma.

2025-01-01T00:00:00Z https://hdl.handle.net/10807/337436 2026-06-02T00:09:30Z 2025-01-01T00:00:00Z

Titolo: Sparse Neural Networks via Bi-level ℓ0 Optimization for Prostate Cancer Prognosis Autori: Frasca, Maria; Lin, Jianyi; La Torre, Davide Abstract: This study proposes a bi-level optimization framework integrating ℓ0 regularization into deep neural networks for joint model compression and embedded feature selection. The approach employs Hard Concrete distributions to approximate the nonconvex ℓ0 norm, enabling simultaneous learning of binary masks for neurons and input features within a differentiable setting. Applied to a cohort of prostate cancer patients, the framework effectively identifies key clinical and molecular predictors while achieving substantial sparsity. Experimental results show 70% neuron pruning and 60% feature reduction with 90% validation accuracy. The method enhances both computational efficiency and clinical interpretability, providing a scalable foundation for decision-support applications in oncology.

2025-01-01T00:00:00Z https://hdl.handle.net/10807/337176 2026-05-30T00:10:19Z 2025-01-01T00:00:00Z

Titolo: Note sulla didattica carducciana Autori: Marco Zanini Abstract: Il presente contributo analizza l'attività di Giosuè Carducci come professore di Letteratura Italiana all'Università di Bologna (1860-1904), focalizzandosi sul suo metodo didattico e sulla sua cruciale influenza sulla formazione della classe docente italiana post-unitaria. Attraverso l'esame di testimonianze dirette di allievi (tra cui Panzini, Albertazzi e Zibordi) e del suo epistolario, l'articolo ricostruisce l'approccio di Carducci, caratterizzato da un forte influsso del Positivismo e del metodo storico-filologico. Si evidenzia come Carducci abbia elevato la storia letteraria a una disciplina scientifica, esigendo la ricerca di dati certi e la rigida analisi del testo, paragonando l'indagine critica a una diagnosi medica o matematica (in analogia con figure come Augusto Murri). L'analisi dimostra che l'impegno di Carducci in cattedra fu primariamente politico-culturale, teso a formare "Italiani compiutamente e consapevolmente" e a sradicare l'estetismo arbitrario in favore della prosa e del rigore. Le note esaminate illustrano come il poeta-docente abbia saputo bilanciare l'alta missione letteraria con una didattica pratica, lasciando un'impronta indelebile sulla critica e sull'organizzazione degli studi umanistici in Italia.

2025-01-01T00:00:00Z