Segmentação automática do pâncreas e massas pancreáticas em tomografias computadorizadas usando arquiteturas encoder-decoder

Abstract

Cancer, the second leading cause of mortality worldwide, accounted for approximately 10 million deaths in 2020. Although relatively rare compared to other malignancies, pancreatic cancer presents one of the poorest prognoses, with a mortality rate of approximately 98%, ranking among the highest across all cancer types. Early-stage detection is the most critical determinant of patient prognosis. However, early diagnosis is particularly challenging due to the difficulty in identifying small lesions in medical imaging modalities such as abdominal ultrasound, computed tomography (CT), and magnetic resonance imaging (MRI). These imaging techniques constitute the primary tools for early detection, and the accurate identification of small pancreatic lesions at early stages significantly improves clinical outcomes. Therefore, there is a pressing need for technologies that can augment and enhance image-based diagnostic workflows. In this work, two deep learning-based methods were developed: one for automatic pancreas segmentation and another for pancreatic mass segmentation in CT scans. The proposed pancreas segmentation approach integrates the EfficientNetB7 backbone within a U-Net architecture, achieving promising results, with a mean Dice similarity coefficient (DSC) of 85.39 ± 2.39% on the NIH dataset and 85.96 ± 2.08% on the MSD dataset. For pancreatic mass segmentation, an ensemble of five encoder-decoder architectures was employed: U-Net, FPN, LinkNet, EDU-Net, and ETDPU-Net. The first three utilize EfficientNetB7 as their encoder backbone. The latter two—EDU-Net and ETDPU-Net—were designed as part of this thesis and incorporate deformable convolutions and Squeeze-and-Excitation (SE) attention mechanisms to enhance feature representation. The outputs of all five models are aggregated using a majority voting ensemble strategy to surpass the performance of individual models. The proposed mass segmentation framework achieved a DSC of 65.28% ± 5.57% on the MSD dataset, validating the efficacy and robustness of the method.