MMPU-Net: A parameter-efficient network for fine-stage of pancreas and pancreas-tumor segmentation on CT scans
Juwita Juwita, Ghulam Mubashar Hassan, Naveed Akhtar, Amitava Datta · Biomedical Signal Processing and Control · 2025
Segmenting organs on CT images is crucial for medical analysis. Recent deep learning methods, such as transformers and U-Net, have been widely used for this task. While transformers focus on global attention and require substantial memory, U-Net-based approaches remain dominant but suffer from several limitations, including spatial information loss due to downsampling, insufficient feature extraction for small and irregular structures, and inefficiency in balancing accuracy with memory constraints. These challenges make it difficult to segment complex organs such as the pancreas and pancreatic tumors, which are often small, occluded, and irregularly shaped. This paper presents three novel deep-learning network designs for pancreas and tumor segmentation, improving accuracy while reducing model size. The key contributions are: (1) replacing traditional downsampling with convolutional strides to better preserve spatial information, (2) integrating MM-block and Point-block in the bottleneck layer to enhance small-structure feature extraction, and (3) employing a hybrid convolution strategy combining standard and depthwise convolutions in both encoder and decoder to optimize accuracy and efficiency. We evaluate the proposed model on NIH and MSD pancreas datasets, using 5188 NIH and 7775 MSD images. MMPU-Net achieves state-of-the-art (SOTA) results, with a Dice Similarity Coefficient (DSC) of 89.53%, Intersection Over Union (IOU) of 81.16%, Precision of 92%, and Recall of 86.75,% on NIH. On MSD, results include 88.60% DSC, 79.90% IOU, 90% Precision, and 87% Recall. For tumor segmentation, MMPU-Net achieves 60.88% DSC, 48.05% IOU, 69.05% Precision, and 61.11% Recall. These contributions highlight MMPU-Net’s effectiveness in balancing efficiency and accuracy for pancreas-tumor segmentation. The code is available at https://github.com/juwita-sj/MMPUNet .