Résumé:
Medical imaging plays a crucial role in the diagnosis and treatment of various diseases,
particularly in identifying and evaluating lesions within the thoracic-abdominal region.
Traditional methods of lesion segmentation, which rely heavily on manual delineation
by radiologists, are time-consuming and prone to variability. Advances in deep learning
and artificial intelligence offer promising solutions to these challenges by automating the
segmentation process, thus enhancing accuracy and efficiency.
This master thesis presents an innovative framework for universal lesion segmentation
in thoracic-abdominal computed tomography (CT) scans using advanced deep learning
techniques. The proposed methodology leverages the U-Mamba model, which integrates
Convolutional Neural Networks (CNNs) and State Space Models (SSMs), to enhance
segmentation accuracy and efficiency.
The framework involves meticulous preprocessing steps, including lesion selection and
Volume of Interest (VOI) cropping, to ensure precise targeting of lesions. The U-Mamba
model, trained with the LION (EvoLved Sign Momentum) optimizer, demonstrates su-
perior performance in capturing long-range dependencies and handling complex lesion
structures across various datasets. The model’s encoder-decoder architecture, coupled
with residual blocks and skip connections, enables robust and detailed segmentation out-
puts. This study validates the efficacy of the proposed framework through comprehensive
evaluations, underscoring its potential to improve diagnostic imaging and clinical work-
flows.
