Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..

Navigating the MIL Trade-Off: Flexible Pooling for Whole Slide Image Classification

Authors: Hossein Jafarinia, Danial Hamdi, Amirhossein Alamdar, Elahe Zahiri, Soroush Vafaie Tabar, Alireza Alipanah, Nahal Mirzaie, Saeed Razavi, Amir Najafi, Mohammad Hossein Rohban

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirically, Maxsoft achieves state-of-the-art performance in low-data regimes across four major benchmarks (CAMELYON16, CAMELYON17, TCGA-Lung, and SICAP-MIL), often matching or surpassing large-scale foundation models. When combined with Per Patch augmentation, this performance is further improved through increased robustness. Code is available at https://github.com/jafarinia/maxsoft
Researcher Affiliation Academia 1Computer Engineering Department, Sharif University of Technology 2Department of Mathematical Sciences, Sharif University of Technology EMAIL EMAIL
Pseudocode Yes Algorithm 1 Multiple Instance Learning of WSI with Maxsoft Algorithm 2 Per Slide Augmentation for WSI Tasks Algorithm 3 Per Patch Augmentation
Open Source Code Yes Code is available at https://github.com/jafarinia/maxsoft
Open Datasets Yes We evaluate on four large-scale pathology WSI datasets CAMELYON16, CAMELYON17, TCGA-Lung and SICAP-MIL [78, 35, 79, 80].
Dataset Splits Yes For CAMELYON16 we use the official split; CAMELYON17, TCGA-Lung and SICAP-MIL are each partitioned into 60% train, 15% val, 25% test (with SICAP-MIL s official split). Details on train/val/test allocations and tiling thresholds appear in Appendix F.
Hardware Specification Yes Experiments use Py Torch 2.1 and scikit-learn on an RTX 4090 [97].
Software Dependencies Yes Experiments use Py Torch 2.1 and scikit-learn on an RTX 4090 [97].
Experiment Setup Yes All models are trained for 500 epochs using the Adam W optimizer [101] with default parameters unless otherwise specified. CAMELYON16. DINO Natural (LR 0.002, WD 0.05, Xavier-uniform); DINO Domain (LR 0.1, WD 0.05, truncated-normal init); UNI (LR 0.02, WD 0.05, Xavier-uniform); Prov-Giga Path (LR 0.02, WD 0.05, Xavier-uniform). CAMELYON17. DINO Natural (LR 0.02, WD 0.005, Xavier-uniform); DINO Domain (LR 0.1, WD 0.05, truncated-normal); UNI (LR 0.02, WD 0.005, Xavier-uniform); Prov-Giga Path (LR 0.02, WD 0.05, Xavier-uniform). TCGA-Lung. DINO Natural (LR 0.02, WD 0.05, Xavier-uniform); DINO Domain (LR 0.002, WD 0.005, truncated-normal); UNI (LR 0.002, WD 0.05, Xavier-uniform); Prov-Giga Path (LR 0.1, WD 0.05, orthogonal). SICAP-MIL. DINO Natural (LR 0.02, WD 0.05, orthogonal); DINO Domain (LR 0.002, WD 0.005, Xavier-uniform); UNI (LR 0.002, WD 0.05, truncated-normal); Prov-Giga Path (LR 0.1, WD 0.05, truncated-normal).