Source-linked AI summary
CAMEL: A Weakly Supervised Learning Framework for Histopathology Image Segmentation
Gang Xu, Zhigang Song, Zhuo Sun, Calvin Ku, Zhe Yang, Cancheng Liu, Shuhao Wang, Jianpeng Ma, Wei Xu
TL;DR
Pixel-level labels for histopathology segmentation are labor-intensive, while image-level supervision alone provides insufficient detail. CAMEL uses MIL-based label enrichment to generate instance labels and approximate pixel masks, achieving segmentation and classification performance comparable with fully supervised approaches. The method may generalize to other weakly supervised histopathology studies, although its direct pixel assignment can over-label regions.
Problem
Fully supervised histopathology segmentation requires labor-intensive pixel-level labels, while weakly supervised methods using only image-level labels lack sufficient supervision information.
Method
CAMEL splits images into latticed instances, uses cMIL to generate instance-level labels, and assigns them to corresponding pixels for segmentation training.
Results
CAMEL achieves comparable performance with fully supervised baselines in both instance-level classification and pixel-level segmentation.
Takeaways & Limitations
CAMEL's automatic labeling methodology may generalize to other weakly supervised histopathology image-analysis studies.
Takeaways & Limitations
Directly assigning instance-level labels to pixels can cause over-labeling, motivating future mask boundary refinement.
Abstract
from arXiv · showhide
Histopathology image analysis plays a critical role in cancer diagnosis and treatment. To automatically segment the cancerous regions, fully supervised segmentation algorithms require labor-intensive and time-consuming labeling at the pixel level. In this research, we propose CAMEL, a weakly supervised learning framework for histopathology image segmentation using only image-level labels. Using multiple instance learning (MIL)-based label enrichment, CAMEL splits the image into latticed instances and automatically generates instance-level labels. After label enrichment, the instance-level labels are further assigned to the corresponding pixels, producing the approximate pixel-level labels and making fully supervised training of segmentation models possible. CAMEL achieves comparable performance with the fully supervised approaches in both instance-level classification and pixel-level segmentation on CAMELYON16 and a colorectal adenoma dataset. Moreover, the generality of the automatic labeling methodology may benefit future weakly supervised learning studies for histopathology image analysis.
1. Introduction
CAMEL addresses the burden and performance limitations of weakly supervised histopathology segmentation by enriching image-level labels into instance-level and approximate pixel-level supervision. Its cMIL-based framework supports fully supervised segmentation training while achieving performance comparable with fully supervised baselines.
- Pixel-level annotation is labor-intensive and time-consuming, motivating weakly supervised histopathology segmentation from coarse image-level labels.
- CAMEL splits images into latticed instances and uses cMIL to automatically generate instance-level labels from image-level labels.The framework comprises label enrichment followed by segmentation.
- The enriched instance labels are assigned to corresponding pixels, producing approximate pixel-level labels for fully supervised segmentation training.
- CAMEL achieves comparable performance with fully supervised baselines in both instance-level classification and pixel-level segmentation, and the colorectal adenoma dataset is released publicly.
- cMIL combines Max-Max and Max-Min instance-selection criteria to balance the data distribution in the constructed instance-level dataset.
- Cascade data enhancement and image-level constraints further use the original supervision information to improve CAMEL.
2. Related Work
Prior weakly supervised histopathology methods face domain-specific morphology challenges and may require artificial constraints, pre-specified features, or complex training procedures. CAMEL instead uses end-to-end cMIL and label enrichment to provide more detailed supervision with less labeling burden.
- Histopathology morphology differs from natural images because foreground and background appearances are less distinct and cancerous regions are disconnected and varied.
- Adversarial erasing and seed-growing methods may be unsuitable for these histopathology characteristics.
- Many MIL approaches require pre-specified features, introducing data-specific prior knowledge that limits general application.
- cMIL performs feature extraction and MIL end-to-end without deliberate feature engineering and feature learning separation.
- CAMEL generates more detailed instance-level supervision with less labeling burden and avoids pre-training requirements compared with CDWS-MIL.
3. Method
CAMEL enriches image-level supervision into instance- and approximate pixel-level labels, then enables fully supervised segmentation training. Its cMIL selection, cascade enhancement, and image-level constraints address label quality, information loss, and supervision use.
- CAMEL framework: CAMEL transforms image-level labels into instance-level labels, which are assigned to corresponding pixels for approximate pixel-level supervision.The enriched labels enable fully supervised training of segmentation models.
- Combined Multiple Instance Learning: cMIL splits each image into N × N latticed instances and uses Max-Max and Max-Min criteria to select instances for a supervised instance-level dataset.Instances from the same image form a bag, and selected instances inherit the image class label.
- Combined Multiple Instance Learning: Max-Max and Max-Min provide complementary selection behavior: Max-Max favors high cancer response, whereas Max-Min counters its tendency to misclassify lower-response cancer instances.For non-cancerous images, selected instances are used to avoid data imbalance.
- Combined Multiple Instance Learning: The two trained classifiers relabel selected instances, combine their outputs, and discard samples whose predicted labels conflict with image-level labels.This produces the final fully supervised instance-level dataset.
- Cascade Data Enhancement: Larger N produces finer enriched labels but causes greater image information loss, motivating cascade enhancement through sequential cMIL stages.The cascade uses cMIL(N1) followed by cMIL(N2), with N = N1 × N2, and combines the resulting data sources.
- Image-Level Constraints: During retraining, CAMEL adds original image-level data through an image-level constraint, defining total loss as weighted retraining and constraint losses.The experiments set the two loss weights equal.
4. Experiments
Experiments evaluate CAMEL’s label enrichment and segmentation on CAMELYON16 and a colorectal adenoma dataset. CAMEL combines complementary instance-selection criteria, image-level constraints, and approximate pixel labels to approach fully supervised performance.
- Data Preparation: CAMELYON16 comprises 400 whole-slide images, with 240 training WSIs and 160 test WSIs used for evaluation.Training data include 5,011 cancer-area and 96,496 non-cancer patches; the test set includes 49 cancer-containing WSIs.
- Label Enrichment: Max-Max tends toward low sensitivity and high specificity, whereas Max-Min tends toward high sensitivity and low specificity.Combining both criteria produces a more balanced trade-off.
- Label Enrichment: Adding cascade data enhancement and image-level constraints further improves CAMEL’s accuracy during instance-level label enrichment.The image-level constraint incorporates original image-level supervision during retraining.
- Label Enrichment: On CAMELYON16, instance classification accuracy is only 1.6% and 1.1% below fully supervised baselines for 320×320 and 160×160 inputs, respectively.The comparisons use the same ResNet-50 classifier architecture.
- Label Enrichment: Instance classification accuracy is 95.5% on 320×320 training sets and 94.6% on 160×160 training sets.These results measure the quality of automatically enriched instance-level labels.
- Segmentation: CAMEL outperforms image-level FSB, WILDCAT, DWS-MIL, and CDWS-MIL, while remaining comparable with pixel-level FSB for segmentation.DeepLabv2 and U-Net are evaluated using sensitivity, specificity, accuracy, and IoU.
- Segmentation: Smaller instance sizes alleviate over-labeling caused by assigning an instance’s cancer label to all its pixels.Finer instances produce finer approximate pixel-level labels and improve segmentation performance.
- Generality: On the colorectal adenoma dataset, CAMEL consistently achieves comparable performance against fully supervised baselines.The dataset contains 177 whole-slide images, including 156 with adenoma.
5. Conclusion
CAMEL uses image-level labels to enrich supervision to the instance level and achieves segmentation results comparable with fully supervised counterparts. The authors suggest that its automatic labeling methodology may generalize to other weakly supervised histopathology studies, while noting over-labeling as a remaining challenge.
- Conclusion: CAMEL automatically enriches supervision from image-level to instance-level labels and achieves segmentation results comparable with fully supervised counterparts.The framework is designed for histopathology image segmentation using only image-level labels.
- Conclusion: The automatic labeling methodology may generalize to other weakly supervised learning studies for histopathology image analysis.This is presented as a potential broader application of CAMEL’s labeling approach.
- Conclusion: Directly assigning instance-level labels to corresponding pixels may produce over-labeling in the segmentation masks.The authors identify mask boundary refinement as future work to address this challenge.