Hossein Jafarinia, Alireza Alipanah, Danial Hamdi, Saeed Razavi, Nahal Mirzaie, Mohammad Hossein Rohban
Whole Slide Image (WSI) classification with multiple instance learning (MIL) in digital pathology faces significant computational challenges. Current methods mostly rely on extensive self-supervised learning (SSL) for satisfactory performance, requiring long training periods and considerable computational resources. At the same time, no pre-training affects performance due to domain shifts from natural images to WSIs. We introduce Snuffy architecture, a novel MIL-pooling method based on sparse transformers that mitigates performance loss with limited pre-training and enables continual few-shot pre-training as a competitive option. Our sparsity pattern is tailored for pathology and is theoretically proven to be a universal approximator with the tightest probabilistic sharp bound on the number of layers for sparse transformers, to date. We demonstrate Snuffy's effectiveness on CAMELYON16 and TCGA Lung cancer datasets, achieving superior WSI and patch-level accuracies. The code is available on https://github.com/jafarinia/snuffy.
| Task | Dataset | Metric | Value | Model |
|---|---|---|---|---|
| Multiple Instance Learning | Elephant | ACC | 0.923 | Snuffy |
| Multiple Instance Learning | Elephant | AUC | 0.967 | Snuffy |
| Multiple Instance Learning | Musk v1 | ACC | 0.961 | Snuffy |
| Multiple Instance Learning | Musk v1 | AUC | 0.989 | Snuffy |
| Multiple Instance Learning | TCGA | ACC | 0.947 | Snuffy (SimCLR Exhaustive) |
| Multiple Instance Learning | TCGA | AUC | 0.972 | Snuffy (SimCLR Exhaustive) |
| Multiple Instance Learning | Musk v2 | ACC | 0.789 | Snuffy |
| Multiple Instance Learning | Musk v2 | AUC | 0.985 | Snuffy |
| Multiple Instance Learning | CAMELYON16 | ACC | 0.948 | Snuffy (DINO Exhaustive) |
| Multiple Instance Learning | CAMELYON16 | AUC | 0.987 | Snuffy (DINO Exhaustive) |
| Multiple Instance Learning | CAMELYON16 | Expected Calibration Error | 0.083 | Snuffy (DINO Exhaustive) |
| Multiple Instance Learning | CAMELYON16 | FROC | 0.675 | Snuffy (DINO Exhaustive) |
| Multiple Instance Learning | CAMELYON16 | Patch AUC | 0.957 | Snuffy (DINO Exhaustive) |
| Multiple Instance Learning | CAMELYON16 | ACC | 0.952 | Snuffy (SimCLR Exhaustive) |
| Multiple Instance Learning | CAMELYON16 | AUC | 0.97 | Snuffy (SimCLR Exhaustive) |
| Multiple Instance Learning | CAMELYON16 | Expected Calibration Error | 0.057 | Snuffy (SimCLR Exhaustive) |
| Multiple Instance Learning | CAMELYON16 | FROC | 0.622 | Snuffy (SimCLR Exhaustive) |
| Multiple Instance Learning | CAMELYON16 | Patch AUC | 0.98 | Snuffy (SimCLR Exhaustive) |
| Multiple Instance Learning | CAMELYON16 | ACC | 0.9 | Snuffy (MAE Adapter) |
| Multiple Instance Learning | CAMELYON16 | AUC | 0.91 | Snuffy (MAE Adapter) |
| Multiple Instance Learning | CAMELYON16 | Expected Calibration Error | 0.078 | Snuffy (MAE Adapter) |
| Multiple Instance Learning | CAMELYON16 | FROC | 0.543 | Snuffy (MAE Adapter) |
| Multiple Instance Learning | CAMELYON16 | Patch AUC | 0.873 | Snuffy (MAE Adapter) |