meetings
A list of DPJC meetings.
2025
- HIBRID: Histology and ct-DNA based risk-stratification with deep learningSrividhya SainathJan 2025
Background: Although surgical resection is the standard therapy for patients with Stage II/III colorectal cancer (CRC) and those with resectable oligometastasis, recurrence rates exceed 30% and 60%, respectively. Circulating tumor DNA (ctDNA) has emerged as a promising predictor of recurrence through the detection of molecular residual disease (MRD). However, morphological tissue information remains unseen by ctDNA. Deep Learning (DL) has been shown to be able to predict prognosis directly from routine histomorphology. By integrating MRD status derived from ctDNA with DL–models trained on histomorphology, we aim to improve patient stratification and recurrence detection. 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 n=3321 patients with resectable stage II-IV CRC. This model was trained on the DACHS cohort (n=1766) and subsequently validated on the GALAXY cohort (n=1555). Patients were categorized into high-risk or low-risk groups based on the DL-prediction scores. In the GALAXY cohort, we combined the DL-scores with the MRD status from the four weeks post-surgery to perform survival analysis. The prognostic efficacy of these biomarkers was assessed using Kaplan-Meier methods and hazard ratios derived from Cox models. Results: In the GALAXY cohort the DL-models categorized 307 patients as high-risk and 1248 patients as low-risk (p<0.001; HR of 2.60, CI 95% 2.11-3.21). ctDNA analysis alone stratified 241 patients as MRD-positive and 1312 patients as MRD-negative (p<0.001; HR of 11.4, CI 95% 9.28-14). Combining these biomarkers, the DFS was significantly stratified in both the MRD-positive group into high-risk (n=81) and low-risk (n=160) with a HR of 1.58 (CI 95% 1.17-2.11; p=0.002) as well as in the MRD negative group into high-risk (n=226) and low-risk (n=1088) with a HR of 2.37 (CI 95% 1.73-3.23; p<0.001). Conclusions: Our results show that even within highly prognostic subsets as yielded by MRD-based stratification, DL is able to further identify patients at high risk of relapse even without measurable MRD. This can be used to improve follow-up and enable early intervention strategies in this vulnerable subset of CRC patients.
- Digital profiling of gene expression from histology images with linearized attentionMarija PizuricaJan 2025
Cancer is a heterogeneous disease requiring costly genetic profiling for better understanding and management. Recent advances in deep learning have enabled cost-effective predictions of genetic alterations from whole slide images (WSIs). While transformers have driven significant progress in non-medical domains, their application to WSIs lags behind due to high model complexity and limited dataset sizes. Here, we introduce SEQUOIA, a linearized transformer model that predicts cancer transcriptomic profiles from WSIs. SEQUOIA is developed using 7584 tumor samples across 16 cancer types, with its generalization capacity validated on two independent cohorts comprising 1368 tumors. Accurately predicted genes are associated with key cancer processes, including inflammatory response, cell cycles and metabolism. Further, we demonstrate the value of SEQUOIA in stratifying the risk of breast cancer recurrence and in resolving spatial gene expression at loco-regional levels. SEQUOIA hence deciphers clinically relevant information from WSIs, opening avenues for personalized cancer management.