[Paper Review] Terabyte-scale Deep Multiple Instance Learning for Classification and Localization in Pathology

Gabriele Campanella, Vitor Werneck Krauss Silva+1 more|arXiv (Cornell University)|May 17, 2018

AI in cancer detection11 references41 citations

TL;DR

The paper presents a terabyte-scale deep MIL framework for whole-slide image classification in prostate needle biopsies using slide-level labels, achieving an AUC of 0.98 on a held-out test set.

ABSTRACT

In the field of computational pathology, the use of decision support systems powered by state-of-the-art deep learning solutions has been hampered by the lack of large labeled datasets. Until recently, studies relied on datasets in the order of few hundreds of slides which are not enough to train a model that can work at scale in the clinic. Here, we have gathered a dataset consisting of 12,160 slides, two orders of magnitude larger than previous datasets in pathology and equivalent to 25 times the pixel count of the entire ImageNet dataset. Given the size of our dataset it is possible for us to train a deep learning model under the Multiple Instance Learning (MIL) assumption where only the overall slide diagnosis is necessary for training, avoiding all the expensive pixel-wise annotations that are usually part of supervised learning approaches. We test our framework on a complex task, that of prostate cancer diagnosis on needle biopsies. We performed a thorough evaluation of the performance of our MIL pipeline under several conditions achieving an AUC of 0.98 on a held-out test set of 1,824 slides. These results open the way for training accurate diagnosis prediction models at scale, laying the foundation for decision support system deployment in the clinic.

Motivation & Objective

• Demonstrate that MIL can scale to terabyte-scale whole-slide images with only slide-level labels.
• Show that large WSI datasets enable high-performance deep MIL models for pathology classification.
• Evaluate how tiling strategies, magnification levels, and class weighting affect MIL performance.
• Identify model architectures and training settings that yield clinically relevant diagnostic accuracy.

Proposed method

• Tile WSIs at multiple magnifications (5x, 10x, 20x) and discard background tiles.
• Treat each slide as a bag of tiles and rank tile-level positive probability within each bag.
• Use the top-1 ranked tile per slide for training with a cross-entropy loss.
• Weight the loss to address class imbalance (weights tested: 0.5, 0.7, 0.9, 0.95, 0.99; final choice w1=0.9).
• Train CNN classifiers (AlexNet, VGG11-BN, ResNet18/34) with Adam optimizer; test with multiple architectures.
• Evaluate by running inference on all tiles at test time and declare slide positive if any tile is positive.

Experimental results

Research questions

• RQ1 Can MIL-based training on terabyte-scale WSIs achieve high diagnostic accuracy using only slide-level labels?
• RQ2 How does dataset size affect MIL performance in whole-slide prostate cancer classification?
• RQ3 Which CNN architectures and magnification strategies maximize MIL performance for WSI diagnosis?
• RQ4 What is the impact of class weighting and data augmentation on MIL generalization?
• RQ5 Do multi-scale ensemble approaches improve MIL performance for WSI classification?

Key findings

• Best models (ResNet34 and VGG11-BN) achieved AUC around 0.976–0.977 on the test set (1,824 slides).
• Overall held-out test AUC reached 0.98 with the top-performing MIL approach.
• False positive rate was 3.7% and false negative rate was 9.4% on the test set after error analysis.
• Ensembling across magnifications (including 5x/10x/20x) improved ROC performance over single magnifications.
• Large dataset size is critical for generalization in MIL-based WSI classification.
• Magnification level affects performance; lower magnifications yielded higher error rates.