In recent years, Deep Learning (DL) techniques have shown strong performance in medical imaging, particularly for automatic pathology classification from Chest X-ray (CXR) images. However, traditional models assume stationary data and often fail in dynamic clinical scenarios, where sequential data shifts can lead to Catastrophic Forgetting (CF), a loss of performance on previously learned tasks. Additionally, DL systems may exhibit biases across demographic groups such as sex and age, raising concerns about fairness in medical decision making. This thesis addresses both CF and demographic bias by exploring multi label pathology classification in a Domain-Incremental Learning (DIL) setting, where CXR images from different healthcare institutions are introduced over time. We implement and evaluate several Continual Learning (CL) strategies, Experience Replay (ER), Dark-Experience Replay (DER), Learning Without Forgetting (LwF), and Pseudo-Labeling, designed to preserve knowledge while adapting to new domains. Fairness is explored through sensitivity-based disparity metrics, monitored throughout training. To contextualize the effectiveness of the proposed methods, we compare them against standard baseline strategies such as fine-tuning and joint training, which represent lower and upper performance bounds, respectively. The objective of this work is to provide a comprehensive assessment of how CL methods affect both predictive performance and fairness in evolving medical environments.
Negli ultimi anni, le tecniche di Deep Learning (DL) hanno ottenuto risultati rilevanti nell’analisi delle immagini radiografiche del torace (CXR), in particolare per la classificazione automatica delle patologie. Tuttavia, i modelli convenzionali assumono che i dati siano stazionari e tendono a fallire in contesti clinici dinamici, dove cambiamenti progressivi nella distribuzione dei dati possono causare il fenomeno del Catastrophic Forgetting (CF), ovvero una perdita di conoscenza su compiti precedentemente appresi. Inoltre, i modelli DL possono manifestare disparità di trattamento tra gruppi demografici, ad esempio in base al sesso o all’età, sollevando questioni legate all’equità nei sistemi di supporto alle decisioni cliniche. Questa tesi affronta congiuntamente il problema del CF e delle disuguaglianze demografiche, analizzando la classificazione multi-label di patologie in uno scenario di Domain-Incremental Learning (DIL), in cui le immagini CXR provenienti da diverse strutture sanitarie vengono introdotte in maniera incrementale. Vengono implementate e confrontate diverse strategie di Continual Learning (CL), tra cui Experience Replay (ER), Dark-Experience Replay (DER), Learning Without Forgetting (LwF) e Pseudo-Labeling, finalizzate a preservare le conoscenze acquisite durante l’adattamento a nuovi domini. L’equità viene monitorata nel tempo tramite metriche basate sulla sensibilità di classe. Per valutare l’efficacia dei metodi proposti, essi vengono confrontati con strategie di riferimento come il fine-tuning e l’addestramento congiunto, che rappresentano rispettivamente i limiti inferiori e superiori delle prestazioni in scenari di apprendimento continuo. L’obiettivo è fornire una valutazione approfondita dell’impatto delle tecniche CL sia in termini di accuratezza predittiva sia di equità in contesti clinici in continua evoluzione.
Exploring Fairness in Domain-Incremental Learning for Chest X-ray Diagnosis
LONGHIN, DILETTA
2024/2025
Abstract
In recent years, Deep Learning (DL) techniques have shown strong performance in medical imaging, particularly for automatic pathology classification from Chest X-ray (CXR) images. However, traditional models assume stationary data and often fail in dynamic clinical scenarios, where sequential data shifts can lead to Catastrophic Forgetting (CF), a loss of performance on previously learned tasks. Additionally, DL systems may exhibit biases across demographic groups such as sex and age, raising concerns about fairness in medical decision making. This thesis addresses both CF and demographic bias by exploring multi label pathology classification in a Domain-Incremental Learning (DIL) setting, where CXR images from different healthcare institutions are introduced over time. We implement and evaluate several Continual Learning (CL) strategies, Experience Replay (ER), Dark-Experience Replay (DER), Learning Without Forgetting (LwF), and Pseudo-Labeling, designed to preserve knowledge while adapting to new domains. Fairness is explored through sensitivity-based disparity metrics, monitored throughout training. To contextualize the effectiveness of the proposed methods, we compare them against standard baseline strategies such as fine-tuning and joint training, which represent lower and upper performance bounds, respectively. The objective of this work is to provide a comprehensive assessment of how CL methods affect both predictive performance and fairness in evolving medical environments.| File | Dimensione | Formato | |
|---|---|---|---|
|
Longhin_Diletta.pdf
Accesso riservato
Dimensione
5.8 MB
Formato
Adobe PDF
|
5.8 MB | 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/20.500.12608/90293