Deep Learning Model to Predict Pancreatic Cancer Outcomes
Author Information
Author(s): Truntzer Caroline, Ouahbi Dina, Huppé Titouan, Rageot David, Ilie Alis, Molimard Chloe, Beltjens Françoise, Bergeron Anthony, Vienot Angelique, Borg Christophe, Monnien Franck, Bibeau Frédéric, Derangère Valentin, Ghiringhelli François, Piergentili Roberto
Primary Institution: Cancer Biology Transfer Platform, Georges-François Leclerc Cancer Centre—Unicancer, Dijon, France
Hypothesis
Can a deep learning approach based on histological images predict the outcomes of pancreatic cancer patients after surgery?
Conclusion
The study shows that deep learning can effectively predict the prognosis of pancreatic cancer and assist in treatment decisions.
Supporting Evidence
- The model identified two patient groups with distinct prognoses.
- There was a significant difference in progression-free survival between the two groups.
- The poor prognosis group was associated with a squamous phenotype.
- The deep learning model improved prediction capacity compared to clinical variables alone.
- High KRT4 expression was linked to a low image-based prognostic score.
Takeaway
Researchers used a computer program to look at images of cancer cells and found that it can help doctors predict how well patients will do after surgery for pancreatic cancer.
Methodology
The study used a deep learning model trained on histological images from a cohort of patients to predict survival outcomes.
Potential Biases
The reliance on retrospective data may introduce biases related to patient selection and treatment variability.
Limitations
The study was retrospective and used heterogeneous cohorts, and the adjuvant treatment primarily consisted of gemcitabine, which is not the current standard of care.
Participant Demographics
The training cohort included 206 patients, with a median age of 67 years, and 50% were male.
Statistical Information
P-Value
0.03
Confidence Interval
[0.54;0.96]
Statistical Significance
p<0.05
Digital Object Identifier (DOI)
Want to read the original?
Access the complete publication on the publisher's website