Avaliação de desempenho de redes neurais profundas para segmentação pulmonar em tomografias computadorizadas do tórax

Abstract

Respiratory diseases rank among the leading causes of death worldwide, making accurate imaging-based diagnostic methods, such as Computed Tomography (CT), essential. In this context, lung semantic segmentation plays a crucial role in the workflow of Computer-Aided Diagnosis (CADe) systems. The objective of this study is to systematically evaluate the performance of different Deep Learning architectures, encompassing both Convolutional Neural Networks (CNNs) and Vision Transformers, in the task of lung segmentation. The experiments were conducted using the LOCCA dataset, composed of volumes from HCU and TASK06. The images were processed through a preprocessing pipeline that included windowing, CLAHE, and median filtering. The architectures UNet++, DeepLabv3+, DPT, and SegFormer were compared using the K-Fold cross-validation technique (k=5) with three repetitions to ensure data robustness. Evaluation metrics included IoU, F1-score, Sensitivity, Specificity, as well as processing time and VRAM consumption. The results indicated that, although the models exhibited statistical equivalence in segmentation quality metrics, there were significant differences in computational cost. The DeepLabv3+ architecture stood out by providing the best balance between performance and resource efficiency, while the DPT model showed the highest time and memory cost.