CECT-Mamba: A Hierarchical Contrast-Enhanced-Aware Model for Pancreatic Tumor Subtyping from Multi-Phase CECT

Zhifang Gong, Shuo Gao, Ben Zhao, Yi Xu, Yijun Yang, Shenghong Ju, GuangQuan Zhou

TL;DR

This paper introduces CECT-Mamba, a novel deep learning model that leverages the Mamba architecture to analyze multi-phase contrast-enhanced CT (CECT) scans for accurate pancreatic tumor subtyping. By effectively capturing both spatial and temporal contrast enhancement patterns across different CECT phases, CECT-Mamba achieves state-of-the-art performance in distinguishing between PDAC and PNET, offering a more efficient and accurate diagnostic tool.

What Problem Does This Paper Address?

Pancreatic tumors are notoriously difficult to diagnose and subtype due to their high heterogeneity and visual similarity, even across different phases of contrast-enhanced computed tomography (CECT). Radiologists rely on subtle changes in how tumors enhance over time (arterial, venous, and delayed phases) to differentiate between subtypes like pancreatic ductal adenocarcinoma (PDAC) and pancreatic neuroendocrine tumors (PNETs). However, this manual interpretation is time-consuming and subjective. Existing automated methods often struggle to fully exploit the rich spatial-temporal information present across these multiple CECT phases, limiting their diagnostic accuracy. This paper aims to bridge this gap by developing an automated system that can effectively learn from and integrate information from all available CECT phases for precise tumor subtyping.

Methods

The core of CECT-Mamba is the Mamba architecture, chosen for its efficiency in modeling long-range dependencies, which is crucial for analyzing sequential data like multi-phase CECT scans. The proposed framework includes several novel components:

  • Dual Hierarchical Contrast-Enhanced-Aware Mamba (DHCM) Module: This module is designed to capture both intra-phase spatial dependencies and inter-phase temporal variations. It uses two complementary sampling sequences: one for spatial modeling that considers global temporal changes, and another for temporal modeling that focuses on local region changes across phases.
  • Similarity-guided Refinement (SimR) Module: Integrated within the temporal modeling, SimR identifies and enhances localized tumor regions that exhibit significant contrast variation across phases, improving the model's focus on diagnostically critical areas.
  • Space Complementary Integrator (SCI) Module: This module aims to mitigate the loss of spatial information that can occur during sequential processing by enriching feature representations with complementary spatial details.
  • Multi-Granularity Fusion (MGF) Module: This component aggregates features from different scales and levels of abstraction, ensuring a comprehensive understanding of the tumor's characteristics.

The overall pipeline involves an encoder-decoder for tumor localization, followed by feature extraction and processing through the DHCM, SCI, and MGF modules before a final classification head.

Datasets

The study utilizes a newly constructed in-house dataset comprising 270 patients with dynamic CECT scans. This dataset includes:

  • 184 patients with PDAC.
  • 96 patients with PNETs.
  • Each patient underwent CECT scans in three phases: arterial, venous, and delayed, resulting in a total of 810 3D images.
  • Tumor regions were annotated by two experienced abdominal radiologists and confirmed via pathological examination.

Key Insights and Ideas

  • Mamba for Medical Imaging Sequences: The paper demonstrates the efficacy of the Mamba architecture, originally developed for sequence modeling, in handling the complex spatial-temporal dynamics of multi-phase medical imaging data. This opens up new avenues for applying Mamba-based models beyond natural language processing.
  • Exploiting Inter-Phase Contrast Dynamics: The core insight is that the subtle, yet critical, changes in contrast enhancement across different CECT phases are key to accurate tumor subtyping. CECT-Mamba's architecture is specifically designed to capture these dynamics.
  • Focusing on Variable Regions: The SimR module's ability to prioritize regions with significant temporal variations highlights that not all parts of a tumor are equally informative for subtyping. This targeted approach improves diagnostic precision.
  • Hierarchical and Multi-Scale Feature Integration: The DHCM, SCI, and MGF modules work synergistically to ensure that both fine-grained local details and broader contextual information across different scales are effectively integrated for robust classification.

Key Results

  • State-of-the-Art Performance: CECT-Mamba achieved an accuracy of 97.41% and an AUC of 98.60% on the in-house dataset, outperforming several leading methods.
  • High Recall and F1-Score: The model also achieved a Recall of 93.13% and an F1-Score of 95.48%, indicating strong performance in correctly identifying positive cases and balancing precision and recall.
  • Computational Efficiency: CECT-Mamba demonstrated competitive inference speed, being the second-fastest among the compared methods, suggesting its potential for clinical deployment.
  • Qualitative Improvement: Grad-CAM visualizations showed that CECT-Mamba focused on more compact and localized regions of the tumor, particularly in challenging cases, aligning with diagnostically critical areas.

Limitations and Future Work

  • Dataset Size and Diversity: While the dataset of 270 patients is substantial, the authors acknowledge that it is an in-house dataset. Future work could involve validating the model on larger, more diverse, multi-center datasets to further assess its generalizability.
  • Clinical Validation: The paper demonstrates strong performance on a retrospective dataset. Further prospective studies and clinical validation would be necessary to confirm its utility in real-world clinical workflows.
  • Generalizability to Other Tumors: The current work focuses on pancreatic tumors. Extending the CECT-Mamba framework to other types of tumors or organs that benefit from multi-phase CECT analysis could be a promising future direction.

Summary updated

Abstract

Contrast-enhanced computed tomography (CECT) is the primary imaging technique that provides valuable spatial-temporal information about lesions, enabling the accurate diagnosis and subclassification of pancreatic tumors. However, the high heterogeneity and variability of pancreatic tumors still pose substantial challenges for precise subtyping diagnosis. Previous methods fail to effectively explore the contextual information across multiple CECT phases commonly used in radiologists' diagnostic workflows, thereby limiting their performance. In this paper, we introduce, for the first time, an automatic way to combine the multi-phase CECT data to discriminate between pancreatic tumor subtypes, among which the key is using Mamba with promising learnability and simplicity to encourage both temporal and spatial modeling from multi-phase CECT. Specifically, we propose a dual hierarchical contrast-enhanced-aware Mamba module incorporating two novel spatial and temporal sampling sequences to explore intra and inter-phase contrast variations of lesions. A similarity-guided refinement module is also imposed into the temporal scanning modeling to emphasize the learning on local tumor regions with more obvious temporal variations. Moreover, we design the space complementary integrator and multi-granularity fusion module to encode and aggregate the semantics across different scales, achieving more efficient learning for subtyping pancreatic tumors. The experimental results on an inhouse dataset of 270 clinical cases achieve an accuracy of 97.4% and an AUC of 98.6% in distinguishing between pancreatic ductal adenocarcinoma (PDAC) and pancreatic neuroendocrine tumors (PNETs), demonstrating its potential as a more accurate and efficient tool.
BibTeX
@misc{gong2025cect,
  author = {Zhifang Gong and Shuo Gao and Ben Zhao and Yi Xu and Yijun Yang and Shenghong Ju and GuangQuan Zhou},
  title = {{CECT-Mamba: A Hierarchical Contrast-Enhanced-Aware Model for Pancreatic Tumor Subtyping from Multi-Phase CECT}},
  year = {2025},
  eprint = {2509.12777},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2509.12777},
}
Figure