Introduction
In the era of precision oncology, Next-Generation Sequencing (NGS) is essential for guiding treatment decisions. However, the success of NGS depends heavily on the initial histopathologic review of Formalin-Fixed Paraffin-Embedded (FFPE) tissue slides. Pathologists must manually estimate tumor purity and determine the number of slides required to achieve a sufficient DNA yield. This "eyeball test" is inherently subjective, often leading to two suboptimal outcomes: under-extraction, which necessitates costly and time-consuming re-processing, or over-extraction, which wastes finite patient tissue.
To address this bottleneck, researchers at Tempus Labs developed SmartPath, an AI-augmented pathology review system. By transitioning from visual estimation to quantitative image analysis, SmartPath serves as a "digital ruler" to optimize DNA yield and preserve precious clinical samples.
The SmartPath System Architecture
SmartPath integrates several advanced computational components into a "pathologist-in-the-loop" workflow:
- Image Segmentation: The system utilizes a U-Net architecture to identify individual nuclei (distinguishing between tumor and immune cells) and a multi-field-of-view convolutional neural network (with a ResNet-18 backbone) to map tumor-rich regions.
- Feature Extraction: The pipeline extracts 3,461 distinct features from the digital slides, including cell counts, nuclear morphology, and spatial texture.
- Interpretable Prediction: Rather than relying on a "black box" model, the researchers used a regularized linear model (L1/Lasso) to correlate these features with historical DNA yield data. This analysis revealed that total cell count is the most significant predictor of yield.
- Decision Support UI: The system provides pathologists with a recommended macrodissection area and a specific slide count required to reach DNA targets (e.g., 100ng, 400ng, or 1000ng).
Clinical Validation and Results
The system was validated through a prospective internal trial involving 501 clinical colorectal cancer (CRC) slides, compared against a traditional manual review cohort. The results demonstrated significant improvements in laboratory efficiency and tissue stewardship:
- Improved Accuracy: The SmartPath-augmented cohort saw a 25% increase in samples falling within the ideal DNA yield range (100–2000ng).
- Tissue Preservation: There was a 14% reduction in "overshooting" (extracting more than 2000ng), ensuring that more tissue remains available for future diagnostic needs.
- Accelerated Turnaround Time: The AI provided the most significant benefit for "small and low-quality" samples. For these difficult cases, SmartPath reduced the time from extraction to sequencing by nearly two days (4.97 days vs. 6.90 days) by avoiding the need for secondary extraction attempts.
Breaking Clinical Habits
One of the study's most interesting findings was the impact of AI on human behavior. Without digital assistance, pathologists tended to default to "round numbers," typically requesting exactly 5 or 10 slides. SmartPath encouraged a more nuanced approach, leading pathologists to request as few as 1–2 slides for high-quality tissue or more than 10 for scant samples, tailoring the extraction to the specific biological reality of the slide.
Limitations and Future Directions
While the trial was successful, the researchers identified real-world variables that can influence outcomes, such as the "weekend effect," where laboratory staffing patterns on specific days can impact turnaround times regardless of AI efficiency. Additionally, while the current model focuses on extrinsic features like cell count, future iterations could incorporate "intrinsic" features—such as DNA fragmentation levels or slide preparation quality—directly from the image to further refine predictions.
Conclusion
SmartPath represents a significant step toward the digitization of the pathology workflow. By providing pathologists with objective, quantitative metrics, the system reduces the variability of manual review, optimizes laboratory resources, and ultimately ensures that cancer patients receive genomic results more quickly and reliably.
Summary updated