DeepLabV3+ and Attention-Based Deep Learning Method for Pancreas Segmentation

The pancreas is one of the body's important organs. Physicians frequently want to visualize the organ's boundaries and shape in radiological images obtained with Computed Tomography (CT) or other imaging technologies. Artificial intelligence-based software has recently gained significant importance to solve such problems. This work presents a method based on the DeepLabV3+ architecture with the ResNet-50 encoder for automatically segmenting pancreatic CT images. A lightweight spatial attention module is integrated into the model's decoder, which improves feature extraction without significantly increasing the computational load by adding only 98 parameters. The model was tested on the NIH Pancreas CT dataset. To address class imbalance in the dataset, a balanced subset was created by sampling from negative images, and contrast enhancement with CLAHE was applied to improve the clarity of the blurred pancreatic borders. Additionally, data augmentation techniques were used to increase the generalizability of the model. Model accuracy was examined using overlap metrics such as Dice Score (DSC) and IoU score, as well as surface distance metrics such as HD95 and ASSD. To demonstrate the superiority of the proposed method, a comparison is made with Attention U-Net and UNET++. As a result, the proposed method achieves high performance with an accuracy of 0.9099 for the DSC metric. The method also achieves high results for other metrics. It outperforms Attention U-Net (0.8792 DSC) by approximately 3.1%.