Atlas-Guided U-Net++ with EfficientNetB5 for Automatic Pancreas Segmentation in Abdominal CT Scans
Felipe R. S. Teles, Neilson P. Ribeiro, Luana Batista da Cruz, Geraldo Bráz, Anselmo Cardoso de Paiva, João Otávio Bandeira Diniz, Omar Andrés Carmona Cortes · 2025
Pancreas segmentation in abdominal computed tomography images is challenging due to the organ’s variability in shape, size, and position. This work proposes an automatic segmentation method based on a 2D Convolutional Neural Network (CNN) approach, consisting of three steps: (1) filtering non-pancreas slices using a CNN, (2) region of interest detection via a probabilistic atlas, and (3) final segmentation with U-Net++ with an EfficientNetB5 backbone. The method achieves a mean Dice coefficient of 78.59% and Recall of 79.12%, with a lower computational cost compared to 2.5 and 3D approaches. Thus, our results stand out among state-of-the-art methods, providing a computationally efficient and accurate solution for diagnosis and treatment planning.