HIBRID: Histology and ct-DNA based Risk-stratification with Deep Learning
Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Background
Although surgical resection is the standard therapy for stage II/III colorectal cancer (CRC), recurrence rates exceed 30%. Circulating tumor DNA (ctDNA) emerged as a promising recurrence predictor, detecting molecular residual disease (MRD). However, spatial information about the tumor and its microenvironment is not directly measured by ctDNA. Deep Learning (DL) can predict prognosis directly from routine histopathology slides.
Methods
We developed a DL pipeline utilizing vision transformers to predict disease-free survival (DFS) based on histological hematoxylin & eosin (H&E) stained whole slide images (WSIs) from patients with resectable stage II-IV CRC. This model was trained on the DACHS cohort (n=1766) and independently validated on the GALAXY cohort (n=1555). Patients were categorized into high- or low-risk groups based on the DL-prediction scores. In the GALAXY cohort, the DL-scores were combined with the four-weeks post-surgery MRD status measured by ctDNA for prognostic stratification.
Results
In GALAXY, the DL-model categorized 307 patients as DL high-risk and 1248 patients as DL low-risk (p<0.001; HR 2.60, CI 95% 2.11-3.21). Combining the DL scores with the MRD status significantly stratified both the MRD-positive group into DL high-risk (n=81) and DL low-risk (n=160) (HR 1.58 (CI 95% 1.17-2.11; p=0.002) and the MRD-negative group into DL high-risk (n=226) and DL low-risk (n=1088) (HR 2.37 CI 95% 1.73-3.23; p<0.001). Moreover, MRD-negative patients had significantly longer DFS when predicted as DL high-risk and treated with ACT (HR 0.48, CI 95% 0.27-0.86; p= 0.01), compared to the MRD-negative patients predicted as DL low-risk (HR=1.14, CI 95% 0.8-1.63; p=0.48).
Conclusion
DL-based spatial assessment of tumor histopathology slides significantly improves the risk stratification provided by MRD alone. Combining histologic information with ctDNA yields the most powerful predictor for disease recurrence to date, with the potential to improve follow-up, withhold adjuvant chemotherapy in low-risk patients and escalate adjuvant chemotherapy in high-risk patients.
Highlights
This study combines MRD status measured by ctDNA with a DL-based risk assessment trained on histological image data to enhance recurrence prediction.
DL-based spatial assessment of tumor histopathology slides significantly improves the risk stratification provided by MRD alone.
MRD-negative patients with high DL-based risk had a significantly longer DFS if treated with ACT, compared to MRD-negative and DL low risk patients
The DL model is fully open-source and publicly available.
