Multiscale Fourier transform multiple instance learning for the Gleason grading of prostate cancer from whole-slide...

Prostate cancer is one of the most prevalent malignancies among men, and accurate diagnosis is crucial for effective treatment planning and prognostication. The Gleason grading system remains the gold standard for histopathological assessment, guiding both diagnosis and risk stratification. However, manual grading is time-consuming and subject to considerable inter-observer variability, which can lead to the misclassification of tumor aggressiveness. Such discrepancies may result in the overtreatment of indolent disease or the undertreatment of aggressive cancer, directly compromising clinical outcomes. Although computational pathology, particularly deep learning, offers a promising solution, its clinical adoption is often hindered by the need for large, pixel-level annotated datasets. Weakly supervised methods such as multiple instance learning (MIL) can address this issue by training on whole-slide image (WSI) labels. However, existing MIL models often struggle to adequately replicate the pathologist's workflow of examining tissue architecture at low magnification and cellular detail at high magnification. This limitation can lead to an incomplete representation of the complex histological landscape and a failure to capture subtle textural patterns crucial for grading. The objective of this study was, therefore, to develop and validate a novel deep learning framework that overcomes these limitations by integrating multiscale spatial information with frequency-domain features. We aimed to create a robust model for automated Gleason grading that relies only on readily available slide-level labels, thereby enhancing diagnostic consistency and efficiency.

To address the limitations of manual pixel-level annotation in Gleason grading, we propose a multiscale Fourier transformer multiple instance learning (MFMIL) model. This approach extracts pathological features at both cellular and glandular levels from multiscale WSIs. A Fourier transform is employed to enhance feature representation and capture subtle morphological variations. A multiscale cross-attention mechanism is then used to integrate features across two magnification levels, effectively reducing semantic discrepancies between resolutions.

The model was validated on two large-scale datasets: 3,775 radical prostatectomy slides from Peking Union Medical College Hospital (PUMCH) and 10,616 biopsy slides from the public Prostate Cancer Grade Assessment (PANDA) dataset. On the PUMCH dataset, MFMIL using only slide-level labels achieved an accuracy of 79.40%, an area under the curve (AUC) of 87.88%, and an average F1-score of 56.54%. On the PANDA dataset, the model achieved an accuracy of 61.97%, an AUC of 85.02%, and an F1-score of 58.46%. Compared to several state-of-the-art baselines (e.g., transformer-based correlated MIL and clustering-constrained attention MIL with multiple branches), MFMIL demonstrated superior performance, particularly in identifying aggressive cases such as grade group 5 (AUC =97.7%).

The proposed MFMIL model demonstrates high diagnostic accuracy and strong clinical applicability for Gleason grading in both radical prostatectomy and prostate biopsy WSIs, offering a promising tool for enhancing prostate cancer diagnosis in routine pathology workflows.