Using Deep Learning to Predict Biochemical Recurrence in Prostate Cancer

Prostate cancer is one of the most commonly diagnosed cancers in men. Many patients undergo radical prostatectomy, a surgical procedure to remove the prostate gland. While this procedure is often successful, about 20% to 40% of patients experience a rise in prostate-specific antigen (PSA) levels after surgery, a condition known as biochemical recurrence. This indicates the possible return of cancer cells and increases the risk of metastasis and mortality.

Predicting which patients are at higher risk of biochemical recurrence before surgery can help doctors tailor treatment strategies. Currently, risk assessment relies on the Gleason grading system, PSA levels, and imaging. However, these methods do not capture subtle histopathological patterns that could provide additional prognostic value. This study aimed to develop a deep learning-based system that analyzes prostate biopsy images to predict the likelihood of biochemical recurrence before surgery.

How the Study Was Conducted

The research team collected prostate biopsy images from 317 patients who underwent radical prostatectomy. Each patient’s dataset included five whole-slide images of tumor tissue. A deep learning model based on the Inception_v3 neural network was trained to analyze the images and predict the likelihood of biochemical recurrence.

The study used a machine learning approach called multiple instance learning, which processes image patches from whole-slide images and integrates them into a predictive model. The deep learning-generated features were combined with clinical data, such as PSA levels and Gleason scores, to improve prediction accuracy. The final model was tested on a separate dataset to evaluate its performance.

Key Findings on Deep Learning Predictions

The deep learning system demonstrated high accuracy in predicting biochemical recurrence, achieving an area under the curve of 0.911 in the testing cohort. This suggests that the model can effectively distinguish between patients who are likely to experience recurrence and those who are not.

Increasing the number of whole-slide images per patient improved prediction accuracy. When all five images were used, the model performed better than when fewer images were analyzed. This highlights the importance of analyzing multiple tumor regions to capture the full complexity of cancer progression.

The study also explored the relationship between deep learning-generated features and traditional pathology findings. The model was able to detect patterns in biopsy images that were associated with higher recurrence risk, reinforcing its potential as a valuable diagnostic tool.

This research demonstrates that deep learning can be used to predict biochemical recurrence in prostate cancer based on biopsy samples. By incorporating this predictive model into clinical workflows, doctors may be able to identify high-risk patients before surgery and adjust treatment plans accordingly.

Patients with a high risk of biochemical recurrence may benefit from additional treatments, such as radiation therapy, hormone therapy, or extended lymph node removal during surgery. Conversely, patients with a low risk of recurrence could potentially avoid overtreatment, reducing unnecessary side effects.

Further studies are needed to validate this model in larger, multi-institutional datasets. Additionally, integrating deep learning predictions with genomic and molecular data could further enhance the accuracy of risk stratification in prostate cancer.

To learn more, read this!: Development of a deep learning system for predicting biochemical recurrence in prostate cancer | BMC Cancer | Full Text