SCIENCE@home

Outcome prediction of prostate cancer patients on active surveillance using weakly supervised deep learning

Filip Winzell,Ida Arvidsson,Karl Åström,Niels Christian Overgaard,F. Marginean,A. Simoulis,Anders Bjartell,A. Krzyzanowska,Anders Heyden

Published 2025 in Medical Imaging

ABSTRACT

To avoid over-treatment of prostate cancer patients following screening for elevated prostate-specific antigen (PSA) levels, keeping patients on active surveillance has been suggested as an alternative to radical treatment. This means reoccurring visits for patients with low-grade cancer to monitor progression. The Prostate Cancer Research International Active Surveillance (PRIAS) study was initiated to research the benefits of active surveillance. Usually, the presence and grade of cancer are determined with needle-core biopsies of the prostate. However, the Gleason grading scale has demonstrated intra- and inter-observer variability. This is a limiting factor for novel automated Gleason grading algorithms. To address this, and simultaneously utilizing a cohort collected within the PRIAS program, we have developed a deep learning-based framework for outcome prediction of patients on active surveillance. Our framework does not use explicit Gleason grades and consists of a pre-trained feature extractor, a feature selector, and an attention-based outcome predictor. We evaluate three feature extractors: the foundation model UNI, an ImageNet pre-trained model, and our Gleason grading network. Using UNI as the feature extractor outperformed the other models, with an average area under the receiver operator characteristic curve (AUC) of 0.996 (95% CI: 0.996 – 0.996). To our knowledge, this is the first end-to-end deep learning-based model for patient-level outcome predictions of prostate cancer patients on active surveillance. We believe that our algorithm could assist the pathologists and facilitate the implementation of prostate cancer screening programs, however, more work is needed in terms of validation and generalization. Our code for training and evaluating our models is publicly available*.

PUBLICATION RECORD

  • Publication year

    2025

  • Venue

    Medical Imaging

  • Publication date

    2025-04-10

  • Fields of study

    Medicine, Computer Science, Engineering

  • Identifiers
    DOI 10.1117/12.3046323
  • External record

    Open on Semantic Scholar

  • Source metadata

    Semantic Scholar

CITATION MAP

EXTRACTION MAP

CLAIMS

  • No claims are published for this paper.

CONCEPTS

  • No concepts are published for this paper.

REFERENCES

Showing 1-37 of 37 references · Page 1 of 1

CITED BY