Abstract A069: Interpretable frequency-domain multiple-instance learning for survival prediction of pancreatic ductal adenocarcinoma
Fabian Cano, Juan Malagón, Diana Barrero, Lyanne Delgado-Coka, Kenneth Shroyer, Eduardo Romero, Luisa Escobar-HoyosAbstract
Background:
Pancreatic ductal adenocarcinoma (PDAC) remains a highly lethal disease, with substantial survival differences among patients with apparently similar clinical categorization. This variability may be related to the histologic heterogeneity of the tumor microenvironment. Whole-slide images (WSIs) contain high-resolution tissue patterns that can provide prognostic and predictive information. However, extracting this information remains challenging because the specific informative regions are not known. Artificial intelligence-based multiple-instance learning (MIL) models can capture relevant information using only case-level labels and can learn which WSI regions are associated with a given outcome. Nevertheless, current attention-based MIL models aggregate data representations as a weighted summation of every instance, filtering out important patterns which are not necessarily the most frequent ones. This limitation was herein addressed by a MIL model which project similarity among instances to the frequency Discrete Cosine (DCT) domain, thereby ensuring attention is isotropically decomposed instead of simply averaged. This formulation sparsely expresses informative tissue regions as a small vector of 64 frequency coefficients per case.
Methods:
A total of 182 cases from the TCGA-PAAD cohort were included, with available overall survival time and clinical information. After tissue quality control and patch extraction, each patch was encoded using the UNI2-h foundation model. A gated-attention MIL module assigned a relevance score to every patch, and the highest-scoring patches from each case were retained. These patches were ordered according to morphological similarity as a structured feature matrix. A two-dimensional DCT was then applied to this matrix, and an 8x8 block of frequency coefficients represented the WSI. The resulting representation was used to train a survival model with a Cox partial-likelihood loss.
Results:
The DCT-based MIL model achieved a test concordance index of 0.67. This value outperformed previously reported image results using TCGA-PAAD, which have been reported as approximately 0.63 in comparable survival prediction settings. Attention scores were mapped back onto the image to generate heatmaps highlighting high-attention regions and identifying areas with patterns potentially related to survival. These results suggest that prognostic information may depend not only on individual high-attention patches, but also on shared frequency-domain patterns across morphologically related tissue regions.
Conclusions:
The proposed model extends conventional attention-based MIL by replacing direct weighted averaging with a frequency-domain representation of jointly selected tissue regions. This approach provides a compact slide-level representation while preserving the spatial origin of the contributing patches.
Citation Format:
Fabian Cano, Juan Malagón, Diana Barrero, Lyanne Delgado-Coka, Kenneth Shroyer, Eduardo Romero, Luisa Escobar-Hoyos. Interpretable frequency-domain multiple-instance learning for survival prediction of pancreatic ductal adenocarcinoma [abstract]. In: Proceedings of the AACR Conference on Pancreatic Cancer: New Frontiers in Biology and Therapeutic Development; 2026 Sep 25-28; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86(18_Suppl_2):Abstract nr A069.