Pancreatic ductal adenocarcinoma (PDAC) staging critically depends on whether the tumor has invaded or is abutting major blood vessels. The Superior Mesenteric Artery (SMA) and Superior Mesenteric Vein (SMV) are the two vessels most relevant for surgical resectability decisions.
When PDAC encases or invades the SMA, the tumor is typically classified as locally advanced and unresectable. Accurate preoperative delineation of tumor boundaries relative to these vessels on CT scans is therefore essential for guiding the decision between upfront surgery, neoadjuvant chemotherapy, or palliative treatment.
Manual segmentation of the pancreatic tumor and adjacent vessels by radiologists is time-consuming, subject to inter-observer variability, and impractical for large-scale screening programs. Automated segmentation tools that accurately delineate both PDAC and the relevant vascular structures would enable standardized, reproducible surgical planning assessments.
The proposed pipeline uses a two-stage approach. In the first stage, a 3D convolutional neural network performs coarse tumor localization, identifying the approximate region of the CT volume containing the pancreas and tumor. This reduces the search space for fine segmentation and avoids the model being distracted by irrelevant anatomy.
In the second stage, the localized region is processed by an Attention U-Net, which extends the standard U-Net architecture with attention gates. Attention gates learn to selectively focus on diagnostically relevant image regions while suppressing background noise, improving segmentation boundary precision in the irregular and heterogeneous context of PDAC.
The two-stage design reflects a general principle in medical image segmentation: coarse-to-fine processing allows each stage to be optimized for its specific sub-task, with the localization stage producing a tight bounding box that focuses the fine segmentation stage on the relevant anatomy.
The key innovation of this paper is TAU-Net (Texture Attention U-Net), which augments the Attention U-Net with hand-crafted texture descriptors. Two texture representation methods are fused with the learned deep features: dense SIFT (Scale-Invariant Feature Transform) and 3D Local Binary Patterns (3D LBP).
Dense SIFT is a classical computer vision descriptor that captures local gradient orientations across multiple scales. Applied densely across the image, it produces rich texture representations that encode the local appearance of tissue in ways complementary to what a CNN learns from supervised training.
3D Local Binary Pattern (3D LBP) encodes the spatial relationship between a voxel and its 3D neighborhood, producing a binary code that summarizes local texture structure. LBP is particularly effective at capturing fine-grained tissue heterogeneity, which is diagnostically relevant because PDAC typically shows irregular, heterogeneous texture compared to normal pancreatic tissue.
The texture descriptor maps are concatenated with the feature maps of the Attention U-Net at multiple decoder levels, allowing the model to leverage both deep learned representations and classical texture descriptions simultaneously. This hybrid approach is motivated by evidence that hand-crafted features capture texture properties that deep networks can underrepresent when trained on small medical imaging datasets.
TAU-Net achieved a Dice Similarity Coefficient (DSC) of 60.6% for PDAC segmentation, representing a 7.52 percentage point improvement over the prior state-of-the-art result on the same dataset. This is a substantial gain for pancreatic tumor segmentation, which is recognized as one of the most challenging medical image segmentation tasks.
A hybrid ensemble model combining predictions from multiple model variants (including TAU-Net with different texture descriptor configurations) further improved performance beyond any single model. Model ensembling reduces variance in predictions and tends to produce more robust segmentation boundaries, particularly for difficult cases.
For vessel segmentation, the model achieved competitive DSC values for both SMA and SMV. Accurate vessel segmentation alongside tumor delineation allows for automated calculation of the tumor-vessel relationship, directly addressing the key surgical planning question of whether PDAC encases or merely abuts these critical structures.
The joint segmentation of PDAC and adjacent vessels is more clinically useful than tumor segmentation alone. Automated outputs that include tumor volume, distance to SMA and SMV, and degree of vascular encasement could be directly integrated into structured radiology reports, providing oncology teams with standardized inputs for staging and treatment planning.
One limitation of current models is that DSC of 60% for PDAC segmentation, while state-of-the-art, still reflects the difficulty of this task. PDAC tumors are often isodense with surrounding pancreatic and peripancreatic tissues on CT, meaning their boundaries have very low contrast. Further performance improvements may require multi-phase CT, multi-modal inputs incorporating PET or MRI, or larger training datasets.
The success of texture descriptor integration in TAU-Net suggests that the field should not abandon classical image analysis approaches in favor of pure deep learning. Hybrid feature architectures that combine learned and hand-crafted representations may be especially valuable in medical imaging, where limited data makes it difficult for deep networks to learn all relevant visual concepts from training examples alone.
The study was conducted on a single-institution dataset, which limits assessment of generalizability. PDAC appearance on CT can vary depending on the CT scanner, acquisition protocol, contrast timing, and patient anatomy, and multicenter validation is needed to confirm that TAU-Net's performance extends across these sources of variation.
Future work could investigate 3D texture descriptors at multiple scales and learned texture representations as alternatives to the fixed SIFT and LBP implementations used here. Neural texture models that are jointly trained with the segmentation task might capture pancreatic tissue texture patterns in ways that pre-defined classical descriptors cannot.
Integration with pre-operative clinical information such as CA 19-9 tumor marker levels, biopsy results, and clinical staging could further improve surgical resectability prediction by combining imaging-based segmentation outputs with non-imaging patient data in a multimodal prediction framework.