Source-linked AI summary
Classification and Disease Localization in Histopathology Using Only Global Labels: A Weakly-Supervised Approach
Pierre Courtiol, Eric W. Tramel, Marc Sanselme, Gilles Wainrib
TL;DR
Histopathology diagnosis requires searching enormous whole-slide images, while pixel-level annotations are costly and often unavailable. CHOWDER uses pretrained features and multiple-instance learning to predict and localize disease from slide-level labels alone. On Camelyon-16, it achieves an AUC of 0.8706 without ground-truth disease maps, and the paper reports performance comparable to top-10 methods trained with strong labels.
Problem
Localized annotations for histopathology are costly and limited, although whole-slide images can contain millions of cells and may have only slide-level diagnoses.
Source: §1 · p. 1Method
CHOWDER adapts multiple-instance learning to classify and localize disease in whole-slide images from slide-level labels, using pretrained tile features and top instances plus negative evidence.
Source: §2.3 · p. 5Results
CHOWDER reaches an AUC of 0.8706 on Camelyon-16 without ground-truth disease maps, and the paper reports performance comparable to top-10 methods trained with strong labels.
Source: §1 · p. 2, §3.3 · p. 7Takeaways & Limitations
The authors conclude that top-instance and negative-evidence learning can support histopathology diagnosis and disease localization without costly expert-produced localized annotations.
Source: §4 · p. 9The reported localization results are preliminary because the study neither samples every tile nor post-processes overlapping tiles into full-resolution maps.
Source: §3.3 · p. 8
Abstract
from arXiv · showhide
Analysis of histopathology slides is a critical step for many diagnoses, and in particular in oncology where it defines the gold standard. In the case of digital histopathological analysis, highly trained pathologists must review vast whole-slide-images of extreme digital resolution ($100,000^2$ pixels) across multiple zoom levels in order to locate abnormal regions of cells, or in some cases single cells, out of millions. The application of deep learning to this problem is hampered not only by small sample sizes, as typical datasets contain only a few hundred samples, but also by the generation of ground-truth localized annotations for training interpretable classification and segmentation models. We propose a method for disease localization in the context of weakly supervised learning, where only image-level labels are available during training. Even without pixel-level annotations, we are able to demonstrate performance comparable with models trained with strong annotations on the Camelyon-16 lymph node metastases detection challenge. We accomplish this through the use of pre-trained deep convolutional networks, feature embedding, as well as learning via top instances and negative evidence, a multiple instance learning technique from the field of semantic segmentation and object detection.
1 INTRODUCTION
CHOWDER addresses the shortage of localized annotations by predicting and localizing disease in whole-slide images from slide-level labels. The approach builds on learned tile features and multiple-instance learning to target histopathology’s large, weakly labeled images.
1 INTRODUCTION
- Localized segmentation maps are costly to produce, leaving annotated histopathology datasets limited and such labels potentially unavailable for new tasks.The paper identifies subtype classification, prognosis estimation, and drug-response prediction as examples of problems where localized annotations may not be available.Source: §1 · p. 1
- Histopathology diagnosis can require many hours of pathologist attention, including searches for a few relevant cells among millions.The introduction notes that diagnostic error rates are difficult to quantify and that false negatives can have serious patient and financial consequences.Source: §1 · p. 1
- Prior histopathology work used handcrafted texture or morphological features with unsupervised processing and conventional classifiers, while deep learning later enabled learned image features.Tile-based methods also used large pretrained networks as feature extractors, avoiding resource-intensive training from scratch.Source: §1 · p. 2
- Whole-slide images span multiple zoom levels and can exceed 100,000 pixels in each dimension, while available diagnoses may label only the slide.With localized annotations difficult to obtain, these datasets fall into weakly supervised learning.Source: §1 · p. 2
- CHOWDER predicts and localizes disease in whole-slide images using only weak, slide-level labels, without expert-produced localized annotations.The paper reports Camelyon-16 performance comparable to top-10 methods trained with pixel-level labels.Source: §1 · p. 2
2 LEARNING WITHOUT LOCAL ANNOTATIONS
The section motivates weakly supervised histopathology analysis because localized annotations are costly, then describes WSI preprocessing, pooled-feature baselines, and CHOWDER’s instance-selection approach. It notes that simple aggregation performs poorly when disease is localized within a slide.
2 LEARNING WITHOUT LOCAL ANNOTATIONS
- Weakly supervised learning seeks suspect WSI regions from image-wide labels, avoiding costly pixel-level annotations; prior work explored EM-based patch selection, decision-level fusion, and WELDON’s top instances with negative evidence.The WELDON framework relaxes the standard multi-instance assumption by considering positive and negative instances.Source: §2 · pp. 2–3
2.1 WSI PRE-PROCESSING
- Preprocessing uses Otsu masks to retain tissue tiles, histogram equalization for color normalization, and fixed-grid non-overlapping tiles to sample slides.The fixed-grid tiles are stored before training; the section motivates this choice by the computational burden of processing very large slides.Source: §2.1 · pp. 3–4
2.2 BASELINE METHOD
- Tiles use ImageNet-pretrained ResNet-50 features; max or mean pooling produces a global slide descriptor for logistic-regression diagnosis.Each tile is represented by 2048 values, and the tile input resolution is 224 × 224 pixels.Source: §2.2 · p. 4, §2.1 · p. 4, Figure 1 · p. 5
2.3 CHOWDER METHOD
- Simple aggregation works for diffuse disease but degrades on Camelyon-16, where localized disease tiles can be overwhelmed by the many healthy tiles.The section contrasts roughly proportional disease and healthy tiles in TCGA-Lung with highly localized disease in Camelyon-16.Source: §2.3 · p. 4
- CHOWDER embeds tile features with a shared one-dimensional convolution; for datasets with fewer than 1000 slides, one embedding is preferred because more can increase generalization error.The embedding uses the same kernel across tiles.Source: §2.3 · p. 5
- CHOWDER retains the top and bottom R embedded tile values as positive and negative evidence, then uses an MLP to classify their interactions.The MinMax selection sends gradients only through the selected tiles during training.Source: §2.3 · pp. 5–6, Figure 2 · p. 5
3 EXPERIMENTAL RESULTS
The experiments evaluate CHOWDER on diffuse TCGA-Lung disease and localized Camelyon-16 metastases, using slide-level classification and lesion-localization measures. On Camelyon-16, CHOWDER performs strongly on classification with only global labels, while localization results are preliminary.
3.1 TRAINING DETAILS
- Training used a fixed 0.5 µm/pixel tile scale, ℓ2 regularization and dropout of 0.5, and Adam optimization; hyperparameters were not tuned.The leaderboard comparison used an ensemble of E = 50 networks with R = 5 and took just over twelve hours to train.Source: §3.1 · p. 6
3.2 TCGA
- For diffuse TCGA-Lung disease, CHOWDER’s advantage over MeanPool is slight, and increasing R to 100 has little apparent effect.The authors suggest global aggregate descriptors can separate the carcinoma classes in this highly homogeneous tissue setting.Source: §3.2 · p. 6
3.3 CAMELYON-16
- The Camelyon-16 evaluation measures slide classification with AUC and metastasis localization with FROC, averaging sensitivity at false-positive rates 0.25, 0.5, 1, 2, 4, and 8.The dataset contains 400 WSIs from sentinel lymph nodes labeled healthy or as containing metastases.Source: §3.3 · pp. 6–7
- Competition-split and random-fold results differ; with MeanPool, the reported AUC is 0.75 on the competition split and 0.55 on random splits.The authors report three-fold cross-validation alongside competition-split results because the distribution mismatch could mislead interpretation.Source: §3.3 · p. 7
- AUC 0.8706 without segmentation masks; CHOWDER outperforms the aggregation baselines and WELDON, while Wang et al. (2016) reached 0.9935 with strong annotations.CHOWDER’s classification AUC is also reported as roughly equivalent to the best-performing pathologist’s 0.884. Optimal performance was obtained with R = 5.Source: §3.3 · p. 7, Table 1 · p. 8, Figure 3 · p. 8
- CHOWDER thresholds the full set of convolutional feature-embedding outputs to classify tiles as diseased or healthy, and its strongest responses occur within the tumor region.The illustrated localization also shows some potential false detections outside the tumor.Source: §3.3 · pp. 7–8, Figure 4 · p. 9
- FROC 0.3103 for CHOWDER in the weak-label setting, versus Wang et al. (2016)’s 0.8074 with strong annotations; the authors call localization results preliminary.They did not fully sample all WSI tiles or post-process overlapping tiles into full-resolution localization maps.Source: §3.3 · p. 8, Figure 3 · p. 8
4 DISCUSSION
The discussion presents CHOWDER as a way to localize disease without costly expert annotations, with potential uses beyond diagnosis. It also notes that localization maps are limited to sampled-tile resolution.
4 DISCUSSION
- CHOWDER uses multiple-instance-learning techniques to predict diagnoses and locate disease in whole-slide images without expert-produced localized annotations.The discussion identifies top instances and negative evidence as the techniques used.Source: §4 · p. 9
- Unannotated localization may support prognosis or drug-response prediction when relevant tissue is not precisely known in advance.The discussion also suggests CHOWDER’s proposed localizations could speed up producing ground-truth localized annotations.Source: §4 · p. 9
- Current localization-map resolution is fixed by the sampled tiles; overlapping tiles with data fusion and heterogeneous annotation are proposed extensions.The heterogeneous setting would combine slides with expert-localized annotations and slides with only whole-slide labels.Source: §4 · p. 9
A FURTHER RESULTS
Figures 5 and 6 visualize CHOWDER’s metastasis detection on two Camelyon-16 test images, pairing whole-slide positive-embedding magnitudes with detailed tile-level classification outputs.
A FURTHER RESULTS
- Figure 5 shows test image 2, with positive feature-embedding magnitudes on the whole slide and tile classification values in a zoomed metastasis detail.Black borders mark ground-truth metastases; classification values run from blue (most negative) through white to red (most positive).Source: Figure 5 · p. 12
- In both figures, uncolored tiles were not included in inference because tiles were randomly selected for inference.Compare the colored tile outputs with the black-bordered ground-truth metastases in each image.Source: Figure 5 · p. 12, Figure 6 · p. 13
- Figure 6 uses the same whole-slide embedding and zoomed tile-classification views for Camelyon-16 test image 92.Positive embedding magnitudes run from white to red, with red indicating the largest magnitude.Source: Figure 6 · p. 13