mpMRI-based habitat analysis for predicting prognoses in patients with high-grade serous ovarian cancer: a multicenter study

Qiu Bi, Kun Miao, Yang Liu, Jing Yang, Ao Zhou, Wenwei Shi, Lei Ying, Yunzhu Wu, Yang Song, Conghui Ai, Haiming Li, Jinwei Qiang

TL;DR

Researchers have developed a "habitat-based" imaging approach that treats ovarian tumors as complex ecosystems rather than uniform masses. By combining spatial sub-regions identified via multiparametric MRI (mpMRI) with standard clinical data, this multicenter study achieved significantly higher accuracy in predicting survival and disease progression for patients with high-grade serous ovarian cancer.

What Problem Does This Paper Address?

High-grade serous ovarian cancer (HGSOC) is the most lethal gynecological malignancy, largely due to its extreme intratumoral heterogeneity. A single tumor isn't a uniform blob; it contains various "neighborhoods" with different blood supplies, cell densities, and genetic profiles.

Standard clinical markers (like FIGO stage or CA125 levels) and conventional "whole-tumor" radiomics often fail to capture this spatial complexity. This paper addresses the need for more precise prognostic tools by using habitat analysis—an unsupervised machine learning technique—to map these internal neighborhoods and see if their specific characteristics can better predict a patient's Overall Survival (OS) and Progression-Free Survival (PFS).

Methods

The study followed a retrospective, multicenter design involving 503 patients. The workflow can be broken down into four key stages:

  1. Habitat Generation: The team used four MRI sequences (T2WI, DWI, ADC, and CE-T1WI). They applied a K-means clustering algorithm to group voxels (3D pixels) with similar signal intensities across these sequences. This divided each tumor into four distinct "habitats" or sub-regions.
  2. Feature Extraction: From these four habitats, they extracted a staggering 27,888 radiomics features (describing shape, texture, and intensity).
  3. Dimensionality Reduction: To avoid overfitting, they used univariate Cox regression followed by LASSO-Cox selection to whittle thousands of features down to the most predictive 45 (for OS) and 62 (for PFS).
  4. Model Building: They constructed three types of models:
  5. Clinical Model: Based on factors like chemotherapy response and platinum resistance.
  6. Habitat Model: Based purely on the MRI sub-region features.
  7. Combined Model: An integration of both clinical and habitat data.

Datasets

The study utilized data from four different medical centers to ensure the model's robustness:

  • Center A (Training & Internal Validation): 314 patients (split 7:3).
  • Centers B, C, and D (External Test Cohort): 189 patients used to verify if the model works on data from different scanners and institutions.

Key Insights and Ideas

  • The "Neighborhood" Matters: The most predictive features weren't spread evenly across the tumor. They were primarily derived from Habitat 2 (characterized by low ADC values and moderate enhancement), which likely represents the dense, solid, and most aggressive components of the tumor.
  • Synergy over Substitution: Interestingly, the habitat-only models were not significantly better than clinical models on their own. However, when combined, the imaging data acted as a powerful "force multiplier," providing a more granular view of the tumor's biology that clinical markers alone couldn't capture.
  • Clinical Anchors: The study confirmed that neoadjuvant chemotherapy (NACT) and platinum resistance remain the strongest clinical predictors of poor outcomes, but the MRI habitats provided the "missing link" to explain why some patients with similar clinical profiles fared worse than others.

Key Results

The Combined Model consistently outperformed both the standalone clinical and habitat models across all metrics:

  • Overall Survival (OS): In the external test cohort, the Combined Model achieved a C-index of 0.745, compared to 0.695 for the clinical model and 0.672 for the habitat model.
  • Progression-Free Survival (PFS): The Combined Model reached a C-index of 0.754 in the external test, significantly higher than the clinical (0.700) or habitat (0.641) versions.
  • 3-Year Risk Prediction: The Area Under the Curve (AUC) for predicting death within three years was 0.821 for the combined model, demonstrating strong potential for identifying high-risk patients early in their treatment journey.

Limitations and Future Work

  • Manual Bottleneck: Tumor segmentation was performed manually. For this to be viable in a high-volume clinical setting, the authors acknowledge that automated, AI-driven segmentation is a necessary next step.
  • Retrospective Nature: As a retrospective study, there is inherent selection bias. The authors suggest that prospective trials are needed to validate these models in real-time clinical decision-making.
  • Imaging Standardization: While the study was multicenter, differences in MRI equipment and protocols across hospitals can introduce "noise" into radiomics data. Future work will focus on standardizing these protocols to improve model generalizability.
  • Biological Grounding: While the habitats correlate with survival, the study did not perform "point-to-point" biopsies to confirm exactly what each habitat represents at a cellular or genetic level (e.g., hypoxia vs. high proliferation).

Summary updated

Abstract

<title>Abstract</title> Purpose To evaluate the value of multiparametric MRI (mpMRI)-based habitat analysis for predicting prognoses in patients with high-grade serous ovarian cancer (HGSOC), and to develop combined models by integrating habitat analysis with clinical predictors. Methods This retrospective study included 503 HGSOC patients from four centers. A K-means algorithm was used to identify voxel clusters and generate habitats on mpMRI. Radiomics features were extracted from each habitat sub-region. After feature selection, habitat models were developed to predict overall survival (OS) and progression-free survival (PFS). Cox regression analyses were performed to identify clinical predictors and construct clinical models. Combined models were developed by integrating habitat signatures with clinical predictors. Model performance was evaluated using C-index and time-dependent receiver operating characteristic area under the curves (AUCs). Results Compared with the clinical models (OS: 0.713 and 0.695; PFS: 0.727 and 0.700) and habitat models (OS: 0.707 and 0.672; PFS: 0.627 and 0.641), the combined models integrating habitat features and clinical independent predictors such as neoadjuvant chemotherapy (OS: 0.752 and 0.745; PFS: 0.784 and 0.754) achieved the highest C-indexes for predicting OS and PFS in the internal validation cohort and external test cohort. The combined models also achieved the highest AUCs in all cohorts. Conclusion The habitat models based on mpMRI demonstrated potential value in predicting the prognoses of HGSOC patients, but no significant advantages over the clinical models. The combined models were expected to improve the prognoses from the level of individual clinical characteristics and habitat features reflecting intratumoral heterogeneity.
BibTeX
@inproceedings{bi2025mpmri,
  author = {Qiu Bi and Kun Miao and Yang Liu and Jing Yang and Ao Zhou and Wenwei Shi and Lei Ying and Yunzhu Wu and Yang Song and Conghui Ai and Haiming Li and Jinwei Qiang},
  title = {{mpMRI-based habitat analysis for predicting prognoses in patients with high-grade serous ovarian cancer: a multicenter study}},
  booktitle = {Abdominal},
  year = {2025},
  doi = {10.1007/s00261-025-05004-9},
}

Loading related content...

Figure