In recent years, the increasingly widespread use of artificial intelligence systems has led to a significant increase in the demand for computational resources, raising questions regarding the energy sustainability and efficiency of the models. This research work fits into this context with the aim of developing and integrating deep neural network architectures that are computationally efficient and suitable for real-world applications in the medical domain, where the scarcity of labelled data and stringent privacy restrictions represent major constraints.Through the project, several directions of research were explored: in a first step, a training process optimisation strategy called Learn\&Drop was proposed, which allows to dynamically identify and remove less relevant layers during training, reducing the number of parameters and improving computational efficiency. On this basis, a lightweight architecture was designed, the Few-Parameter Architecture (FPA), capable of maintaining competitive performance in classification and segmentation while operating with reduced computational resources. The experimental validation of these models was conducted on real biomedical data, characterised by high variability and the presence of artefacts, in order to meet realistic and challenging requirements aligned with the state of the art.A primary issue that emerged lies in the integration of AI into clinical flows, which led to the development of IODeep, a DICOM-compliant data structure that enables complaint entry into hospital PACS, ensuring patient privacy and compliance with field protocols. The domain considered presents an additional challenge, namely the scarcity of annotated data, which in the first instance was addressed using generative AI approach to generate cross-domain synthetic images useful to fill this gap. In addition, a meta-learning-based approach based on the few-shot segmentation paradigm was developed. In this way, even where synthetic generation is not sufficiently reliable, it is possible to provide clinical practice with tools capable of operating under conditions of low data availability.Overall, the results obtained highlight how it is possible to reconcile predictive accuracy, computational efficiency and environmental sustainability through the adoption of innovative architectural design techniques and adaptive learning strategies, thus contributing to the development of viable and safe artificial intelligence solutions in the health sector.
Negli ultimi anni, l’utilizzo sempre più diffuso di sistemi di intelligenza artificiale ha comportato un aumento significativo della domanda di risorse computazionali, sollevando interrogativi in merito alla sostenibilità energetica e all’efficienza dei modelli. Questo lavoro di ricerca si inserisce in tale contesto con l’obiettivo di sviluppare e integrare architetture di rete neurale profonda che siano computazionalmente efficienti e adatte ad applicazioni reali nel dominio medicale, dove la scarsità di dati etichettati e le rigorose restrizioni legate alla privacy rappresentano vincoli rilevanti.Nel corso del progetto, sono state esplorate diverse linee di ricerca: in una prima fase è stata proposta una strategia di ottimizzazione del processo di addestramento denominata Learn&Drop, che permette di identificare e rimuovere dinamicamente i layer meno rilevanti durante l’addestramento, riducendo il numero di parametri e migliorando l’efficienza computazionale. Su questa base, è stata progettata un’architettura leggera, la Few-Parameter Architecture (FPA), capace di mantenere prestazioni competitive in classificazione e segmentazione, pur operando con risorse computazionali ridotte. La validazione sperimentale di questi modelli è stata condotta su dati biomedicali reali, caratterizzati da elevata variabilità e presenza di artefatti, al fine di rispondere a requisiti realistici e sfidanti, allineati con lo stato dell’arte. Una delle prime problematiche emerse risiede nell’integrazione dell’Intelligenza Artificiale nei flussi clinici, ragion per cui è stata sviluppato IODeep, una struttura dati conforme allo standard DICOM, che consente l’inserimento complaint nei PACS ospedalieri, garantendo la privacy dei pazienti e l’aderenza alle normative di settore. Il dominio considerato presenta un ulteriore problematica, ovvero la scarsità di dati annotati che in prima istanza è stato affrontato utilizzando approccio di AI generativa, per generare immagini sintetiche cross-domain utili a sopperire questa carenza. In aggiunta, un approccio basato sul paradigma del few-shot segmentation, di meta-learning-based, è stato sviluppato. In questo modo, anche laddove la generazione sintetica non risulti sufficientemente affidabile, è possibile fornire alla pratica clinica strumenti in grado di operare in condizioni di bassa disponibilità di dati.Nel complesso, i risultati ottenuti evidenziano come sia possibile conciliare accuratezza predittiva, efficienza computazionale e sostenibilità ambientale attraverso l’adozione di tecniche innovative di progettazione architetturale e strategie di apprendimento adattive, contribuendo così allo sviluppo di soluzioni di intelligenza artificiale praticabili e sicure nel settore della salute.
Developing and integrating computationally efficient deep neural networks for medical image segmentation
CRUCIATA, Luca
2025
Abstract
In recent years, the increasingly widespread use of artificial intelligence systems has led to a significant increase in the demand for computational resources, raising questions regarding the energy sustainability and efficiency of the models. This research work fits into this context with the aim of developing and integrating deep neural network architectures that are computationally efficient and suitable for real-world applications in the medical domain, where the scarcity of labelled data and stringent privacy restrictions represent major constraints.Through the project, several directions of research were explored: in a first step, a training process optimisation strategy called Learn\&Drop was proposed, which allows to dynamically identify and remove less relevant layers during training, reducing the number of parameters and improving computational efficiency. On this basis, a lightweight architecture was designed, the Few-Parameter Architecture (FPA), capable of maintaining competitive performance in classification and segmentation while operating with reduced computational resources. The experimental validation of these models was conducted on real biomedical data, characterised by high variability and the presence of artefacts, in order to meet realistic and challenging requirements aligned with the state of the art.A primary issue that emerged lies in the integration of AI into clinical flows, which led to the development of IODeep, a DICOM-compliant data structure that enables complaint entry into hospital PACS, ensuring patient privacy and compliance with field protocols. The domain considered presents an additional challenge, namely the scarcity of annotated data, which in the first instance was addressed using generative AI approach to generate cross-domain synthetic images useful to fill this gap. In addition, a meta-learning-based approach based on the few-shot segmentation paradigm was developed. In this way, even where synthetic generation is not sufficiently reliable, it is possible to provide clinical practice with tools capable of operating under conditions of low data availability.Overall, the results obtained highlight how it is possible to reconcile predictive accuracy, computational efficiency and environmental sustainability through the adoption of innovative architectural design techniques and adaptive learning strategies, thus contributing to the development of viable and safe artificial intelligence solutions in the health sector.File | Dimensione | Formato | |
---|---|---|---|
Tesi_Cruciata_1507.pdf
accesso aperto
Dimensione
18.03 MB
Formato
Adobe PDF
|
18.03 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/20.500.14242/215217
URN:NBN:IT:UNIPA-215217