Optimizing lung cancer understanding: leveraging active learning for data-efficient AI

Lung cancer remains the leading cause of cancer-related mortality worldwide, yet AI- driven prognoses are often hampered by the high cost and limited availability of richly annotated clinical data. In this thesis, we develop a data-efficient Active Learning frame- work centered on a Multi-Layer Perc...

ver descrição completa

Detalhes bibliográficos
Autor: Kokin, Danila
Formato: tesis de maestría
Fecha de publicación:2025
País:España
Recursos:Universitat Politècnica de Catalunya (UPC)
Repositorio:UPCommons. Portal del coneixement obert de la UPC
Idioma:inglés
OAI Identifier:oai:upcommons.upc.edu:2117/449758
Acesso em linha:https://hdl.handle.net/2117/449758
Access Level:acceso abierto
Palavra-chave:Lungs--Cancer
Artificial intelligence--Medical applications
Active learning
Tabular data
Medical data
Lung cancer
Bayesian neural network approximation
Pulmons--Càncer
Intel·ligència artificial--Aplicacions a la medicina
Àrees temàtiques de la UPC::Informàtica::Intel·ligència artificial
Descrição
Resumo:Lung cancer remains the leading cause of cancer-related mortality worldwide, yet AI- driven prognoses are often hampered by the high cost and limited availability of richly annotated clinical data. In this thesis, we develop a data-efficient Active Learning frame- work centered on a Multi-Layer Perceptron enhanced with Monte Carlo Dropout to ap- proximate Bayesian uncertainty on structured patient records. Using a retrospective cohort from the Hospital of Barcelona-encompassing demographic, metastatic, labora- tory, histological, and molecular biomarker information-our model begins with a small seed of labeled cases. At each iteration, we perform multiple stochastic forward passes to estimate predictive variance and then query the most uncertain samples for expert anno- tation. We systematically compare class-balancing techniques (SMOTE, ADASYN, and weighted losses) and a variety of sampling strategies-including random, cluster-based, uncertainty-driven, distance-based, and hybrid approaches-across different seed-pool and batch-size configurations. Focused on treatment-response classification, this strategy matches or exceeds the performance of fully supervised models while reducing labeling requirements by up to 60%. We conclude by discussing the trade-offs between annota- tion cost, model confidence, and clinical utility, highlighting Active Learning's potential to accelerate AI adoption in data-constrained healthcare environments.