Introduction
High-grade serous ovarian cancer (HGSOC) is characterized by its aggressive spread throughout the abdominal cavity, primarily affecting the pelvic region, ovaries, and the omentum. In clinical practice, treatment response is typically monitored using the RECIST 1.1 criteria, which rely on simplified one- or two-dimensional measurements of specific "target" lesions. While these metrics are standard, they often fail to capture the full volumetric complexity of the disease.
Volumetric segmentation—mapping the entire three-dimensional tumor burden—offers a more comprehensive assessment of chemotherapy response. However, manual segmentation is labor-intensive and time-consuming for radiologists. This study presents a fully automated deep learning-based method designed to segment multi-site ovarian cancer lesions on computed tomography (CT) scans, aiming to bridge the gap between clinical precision and workflow efficiency.
Methodology
Model Architecture and Optimization
The researchers developed their model by building upon the nnU-Net (no-new-Net) framework, the current benchmark for medical image segmentation. The team implemented several strategic modifications to enhance performance for the specific challenges of ovarian cancer imaging:
- Architectural Refinement: The standard six-stage U-Net was streamlined to four stages, and the encoder was replaced with a ResNet backbone (utilizing 1, 2, 6, and 3 blocks per stage).
- Hyper-parameter Tuning: A core finding of the study was that "careful tuning" often outweighed radical architectural changes. The team optimized the training regime by increasing batch sizes, adjusting weight decay, and employing a sophisticated learning rate schedule featuring linear ascent followed by cosine decay.
- Progressive Learning: To improve stability, the model was trained using a progressive strategy, starting with lower-resolution patches and simpler data augmentations before gradually increasing complexity.
Dataset and Multi-site Validation
The study utilized a diverse dataset of 451 CT scans sourced from four institutions across the United Kingdom and the United States. This multi-site approach ensured the model was exposed to various scanner manufacturers (including Siemens, GE, Philips, and Toshiba) and differing clinical protocols. The data included both pre-treatment and post-neoadjuvant chemotherapy (NACT) scans, providing a realistic representation of tumor morphology at different stages of treatment.
Results and Clinical Comparison
The model's performance was evaluated using the Dice Similarity Coefficient (DSC), where a score of 1.0 indicates perfect overlap with ground-truth manual segmentations.
Performance Benchmarks
- Pelvic/Ovarian Lesions: The model achieved a mean DSC of 72 ± 19 on the independent test set. Most significantly, statistical analysis (p=0.392) showed no significant difference between the model’s performance and that of a trainee radiologist with three years of specialized experience.
- Omental Lesions: These lesions proved more challenging due to their often thin, "smudgy" appearance, resulting in a mean DSC of 64 ± 24.
- Baseline Comparison: The tuned model significantly outperformed the standard nnU-Net framework across all test sets (p < 0.005), validating the importance of the researchers' optimization strategies.
Discussion and Future Directions
The study demonstrates that automated segmentation of ovarian cancer is not only feasible but can reach a level of accuracy comparable to human trainees for pelvic disease. This suggests that AI-driven volumetric assessment could soon become a viable tool for tracking total tumor burden in clinical trials and routine practice.
Limitations
Despite its success, the model faced specific hurdles. Because it processes data in localized "patches," it occasionally generated false positives in non-abdominal regions like the thighs or breast tissue. Additionally, "conglomerate" tumors—where pelvic and omental diseases merge—presented difficulties in boundary definition.
Future Work
To address these limitations, the authors suggest integrating Vision Transformers (ViTs) to better capture long-range anatomical relationships and implementing global organ segmentation maps to mask out irrelevant anatomical regions. Expanding the model to detect less common metastatic sites remains a priority for achieving a truly comprehensive automated assessment of HGSOC.
Summary updated