TL;DR
This paper introduces a novel method for precisely comparing PET/CT scans with detailed prostate cancer pathology at a voxel-by-voxel level. By creating 3D "histoPET" models from histopathology and using advanced registration techniques, they achieved a more accurate understanding of how PET signals correlate with actual tumor locations, paving the way for better treatment planning.
What Problem Does This Paper Address?
Diagnosing and treating prostate cancer (PCa) with PET/CT requires a deep understanding of how the imaging tracer behaves within the prostate, both in cancerous and healthy tissue. However, directly comparing PET/CT images (which have millimeter-scale resolution) with detailed histopathology slides (which have micron-scale resolution) is incredibly challenging. This difficulty stems from several factors:
- Resolution Mismatch: PET images show larger structures, while pathology reveals fine cellular details.
- Deformation: The prostate can change shape significantly during surgery and processing for pathology.
- Image Modalities: PET and CT are fundamentally different types of images.
- Registration Uncertainty: Accurately aligning these different types of data (coregistration) is crucial but prone to errors, especially when relying solely on CT.
Previous studies often relied on visual comparisons or simpler registration methods, limiting the precision of their findings. This paper aims to overcome these limitations by developing a more robust, quantitative, voxel-wise comparison strategy.
Methods
The core of this study lies in a multi-step process to achieve precise voxel-wise comparison:
- Histopathology Preparation and 3D Modeling:
- Prostate tissue from surgery was meticulously prepared and stained.
- Each slice was scanned and its histopathological information (including tumor boundaries) was digitized.
- This 2D information was used to construct a 3D "histoPET" model. Crucially, this model was then smoothed to mimic the lower spatial resolution of PET scans, creating a "relSUV" (relative Standardized Uptake Value) that represents a smoothed tumor likelihood.
- Multi-Stage Coregistration:
- CT-based Coregistration: The in vivo (pre-surgery) CT scan and the ex vivo (post-surgery) CT scan of the prostate were aligned using three different methods:
- Manual Registration (ManReg): Expert observers manually aligned the scans.
- Manual Registration with Automatic Scaling (ScalFactReg): Manual alignment was combined with an automatic scaling factor to account for prostate shrinkage.
- Deformable Registration (DefReg): An advanced algorithm was used to account for non-linear deformations.
- PET-based Mutual Information (MI) Coregistration: After CT-based alignment, a rigid mutual information algorithm was applied to align the histoPET model with the actual PSMA-PET/CT scan. This step leverages the intensity patterns in both images to find the best alignment.
- Voxel-wise Analysis:
- Once aligned, the correlation between PSMA-PET signal (SUV) and the histoPET signal (relSUV) was quantified using the coefficient of determination (R²). This measures how well the PET signal can be predicted by the histopathology.
Datasets
The study utilized data from eight patients who underwent radical prostatectomy for histopathologically proven primary prostate cancer. Pre-operative PSMA-PET/CT scans were acquired, and the resected prostates were processed for detailed histopathology. No new benchmark datasets were introduced; rather, existing patient data was used to develop and validate the proposed methodology.
Key Insights and Ideas
- The Power of 3D Modeling: Creating a 3D histoPET model from detailed pathology, and then smoothing it to match PET resolution, is key to enabling meaningful voxel-wise comparisons. It bridges the gap between microscopic and macroscopic views.
- CT-based Registration is a Necessary First Step: While not perfect, CT-based registration (especially deformable methods) provides a crucial anatomical starting point for aligning the in vivo and ex vivo data.
- PET-based MI is Essential for Precision: Adding a mutual information coregistration step after CT-based alignment significantly improves the correlation between PET and histopathology. This suggests that relying solely on anatomical landmarks from CT is insufficient.
- Manual Registration + MI is a Strong Contender: The combination of manual CT-based registration (ManReg) followed by PET-based MI (ManRegMI) offered a robust and precise alignment, often outperforming more complex automated methods in terms of correlation strength.
- Resolution Matters: The study explicitly acknowledges that small tumors missed in histopathology could lead to apparent "false positives" in PET. This highlights the need for comprehensive histopathology and careful interpretation.
Key Results
- CT-based Registration Performance: Deformable registration (DefReg) achieved the highest Dice Similarity Coefficient (DSC), indicating the best spatial overlap of prostate contours between in vivo and ex vivo CT scans (mean DSC of 0.90 ± 0.02).
- Voxel-wise Correlation (R²):
- After CT-based registration alone, R² values were moderate, ranging from 11.4% to 67.9%.
- After the subsequent PET-based MI coregistration, R² values significantly increased, reaching up to 80.7% for ManRegMI, 79.1% for ScalRegMI, and 79.1% for DefRegMI. This demonstrates that the MI step substantially improved the agreement between PET and histopathology.
- MI Improves Agreement: The MI step consistently led to higher R² values across all CT-based registration methods, indicating a better quantitative match between the PSMA-PET signal and the histopathology-derived histoPET model.
Limitations and Future Work
- Histopathology Coverage: The study acknowledges that the histopathology coverage of the entire prostate volume is technically demanding and can be incomplete. This missing information can impact the accuracy of the histoPET model and the resulting correlations. Future work could focus on improving the comprehensiveness of histopathological preparation.
- Deformable Registration Sensitivity: The deformable registration method (DefReg) was found to be sensitive to contour uncertainties, potentially leading to pronounced deformations. This suggests that while powerful, it requires very accurate input data.
- Manual Registration Stability: While manual registration (ManReg) can be labor-intensive, it appeared to be more stable and less susceptible to certain artifacts compared to automated methods in this study.
- Future Directions: The authors recommend their combined manual CT-based and PET-based MI coregistration approach as a less complex yet stable and precise method for voxel-wise comparisons. This methodology is seen as a crucial step towards enabling advanced applications like "dose-painting by numbers" in focal radiotherapy, where precise knowledge of tumor location and extent is paramount.
Summary updated