Research from Lund University in Sweden showed that a new AI framework achieved high accuracy in predicting patient outcomes without relying on traditional Gleason grading methods. This approach extends beyond slide-level predictions to patient-level outcomes in the WSI setting.

Filip Winzell, doctoral researcher at Lund University, and his colleagues developed a framework that predicts whether a patient should remain on active surveillance or receive immediate treatment, achieving an area under the curve 0.996.[1] Rather than focusing on Gleason patterns, the system analyzes WSIs to make patient-level predictions.

Winzell explained that although prostate-specific antigen (PSA) screening can reduce mortality, it leads to unacceptable levels of overtreatment in some patients. “Active surveillance has been proposed by several studies to reduce the risk of overtreatment following PSA screening, but population-wide screening program would increase the workload for pathologists,” Winzell said.

The traditional approach to prostate cancer diagnosis involves examining 6–12 biopsies per patient, with each generating approximately one gigabyte of WSI data. Pathologists must manually grade these images using the Gleason system, which is a time-consuming process with high inter- and intra-observer variability.

The team developed a framework trained to predict patient outcomes (treatment decisions made by pathologists) rather than calculating Gleason grades. When asked about the advantages of bypassing Gleason grading, Winzell said: “Although patient outcomes are correlated with Gleason grades, they are big-picture assessments, where the pathologists can weigh in many more factors. With this objective, the model will learn from several treatment decisions and could thereby provide a strong second opinion in a clinical setting.”

The pipeline begins with tissue segmentation and patch extraction. A feature extractor then processes these patches, followed by feature selection and attention-based learning to generate the final prediction. The team compared three feature extractors: their own GG Net (pre-trained to identify malignant patches), the UNI foundation model, and an ImageNet-based approach.

Winzell explained that the feature extractor serves two crucial purposes: “First, it represents patches with feature vectors, processing the vast information in the WSIs while preserving relevant details. Second, it helps overcome our limited dataset size by leveraging pre-training on external data.”

Winzell added that the attention-based learning mechanism of the system allows all instances to contribute to the prediction and enables the model to weigh which are more important. “Without it, we would need an enormous dataset to reach the same level of performance,” he noted.

The research team validated their approach using data from the Prostate Cancer Research International Active Surveillance study, analyzing 145 patients with approximately 3,000 images. The UNI-based model demonstrated the best performance, significantly outperforming both the ImageNet approach and traditional protocol-based methods.

During his talk, Winzell shared that the system’s attention patterns revealed some unexpected findings. “For treated cases, we observed that sometimes the model rated even benign patches higher than some Grade 4 patterns,” he said. “This suggests there might be prognostic information present even in seemingly benign tissue areas.”

Regarding the potential clinical implications of this work, Winzell stated that the framework could mitigate some of the issues of current screening programs and prevent potential overtreatment following PSA screening by making accurate predictions. “Our framework could be useful as a second opinion for difficult cases, as well as filtering out easy, clear cases,” he added.

However, Winzell acknowledges potential barriers to the clinical implementation of their framework. “The main challenge for clinical implementation is that it would require a much more thorough analysis of its generalizability,” he cautioned. “Our dataset is quite homogeneous in terms of staining and slide scanner use. Gathering this type of data requires following patients and collecting data over several years, which is challenging.”

In the future, the team will explore ways to further enhance the system. “We have considered including PSA measurements and prostate volume in our framework,” Winzell revealed. “Initial results were somewhat inconclusive, but it’s something worth studying more.”Filip Winzell, doctoral researcher at Lund University, Sweden.

[1] Filip Winzell, Outcome prediction of prostate cancer patients on active surveillance using weakly supervised deep learning. Presented at SPIE 2025 Digital and Computational Pathology conference, February 19, 2025; San Diego, CA.