Source-linked AI summary
Deep Learning for Identifying Metastatic Breast Cancer
Dayong Wang, Aditya Khosla, Rishab Gargeya, Humayun Irshad, Andrew H. Beck
TL;DR
The paper addresses automated detection of metastatic breast cancer in sentinel lymph node whole slide images, where manual review is burdensome and small metastases are difficult to identify. It uses deep patch classification with heatmap-based post-processing for slide classification and lesion localization. The system won both Camelyon16 competitions, and combining its predictions with a pathologist reduced the pathologist’s error rate from over 3 percent to less than 1 percent.
Problem
Manual sentinel lymph node review is time-consuming and difficult for small metastatic foci, while existing computer-assisted systems were not used clinically.
Method
A deep convolutional network classifies millions of tissue patches, whose predictions are aggregated into heatmaps for slide classification and lesion localization.
Results
The system won both Camelyon16 competitions, achieving 0.9250 slide-level AUC and a 0.7051 lesion-based score.
Takeaways & Limitations
Combining deep learning predictions with pathologist interpretations reduced pathologist error rates from over 3 percent to less than 1 percent.
Abstract
from arXiv · showhide
The International Symposium on Biomedical Imaging (ISBI) held a grand challenge to evaluate computational systems for the automated detection of metastatic breast cancer in whole slide images of sentinel lymph node biopsies. Our team won both competitions in the grand challenge, obtaining an area under the receiver operating curve (AUC) of 0.925 for the task of whole slide image classification and a score of 0.7051 for the tumor localization task. A pathologist independently reviewed the same images, obtaining a whole slide image classification AUC of 0.966 and a tumor localization score of 0.733. Combining our deep learning system's predictions with the human pathologist's diagnoses increased the pathologist's AUC to 0.995, representing an approximately 85 percent reduction in human error rate. These results demonstrate the power of using deep learning to produce significant improvements in the accuracy of pathological diagnoses.
1. Introduction
Pathology requires accurate, reproducible diagnoses, but manual microscopic review is difficult and burdensome. The paper presents a deep learning system for detecting metastatic breast cancer in sentinel lymph node whole slide images and reports strong competition performance with pathologist integration.
- Manual microscopic pathology review can lack standardization, produce diagnostic errors, and impose substantial cognitive load.Pathologists may need to evaluate millions of cells across hundreds of slides in a workday.
- Camelyon16 evaluated computational systems for automatically detecting metastatic breast cancer in sentinel lymph node whole slide images.Sentinel lymph node status affects breast cancer staging and can lead to more aggressive clinical management.
- Manual sentinel lymph node review is time-consuming, especially for cancer-negative nodes or small metastatic foci.Pancytokeratin immunohistochemistry can improve sensitivity but increases cost, preparation time, and slide burden.
- Existing computer-assisted systems for detecting small metastatic foci were not used clinically, motivating effective and cost-efficient evaluation methods.A high-performing system could increase accuracy and reduce cognitive load at low cost.
- The proposed system trains a deep convolutional network on millions of patches, aggregates predictions into heatmaps, and supports slide classification and tumor localization.Combining system predictions with pathologist interpretations significantly reduced the pathologist’s error rate.
2. Dataset and Evaluation Metrics
Camelyon16 supplied whole slide images and metastatic-region annotations for training and testing. Participants were evaluated separately on slide-level discrimination and lesion-level detection using AUC and average sensitivity across predefined false-positive rates.
- 400 whole slide images comprised the Camelyon16 dataset, with 270 for training and 130 for testing.Both splits contained samples from Radboud UMC and UMC Utrecht.
- Training ground truth consisted of pathologist-delineated metastatic regions represented as XML contours and binary masks.
- Slide-based evaluation measured discrimination between metastatic and normal slides using area under the receiver operating characteristic curve.Participants submitted one cancer probability for each test slide.
- Lesion-based evaluation measured average sensitivity for detecting true cancer lesions across six false-positive rates per whole slide image.The predefined rates were 1/4, 1/2, 1, 2, 4, and 8 false positives per whole slide image.
3. Method
The method detects tissue, classifies image patches with deep networks, converts predictions into tumor-probability heatmaps, and post-processes those heatmaps for slide classification and lesion localization.
- 3.1. Image Pre-processing: Tissue detection uses HSV conversion and Otsu thresholds to exclude background white space before analysis.The resulting masks combine thresholded H and S channels.
- The framework comprises a patch-based classification stage followed by heatmap-based post-processing.
- 3.2. Cancer Metastasis Detection Framework: Millions of positive and negative patches are extracted from annotated whole slide images, classified, and embedded into tumor-probability heatmaps.Positive patches are tumor regions labeled 1, while negative patches are normal regions labeled 0.
- 3.3. Patch-based Classification Stage: GoogLeNet was selected for patch classification after deeper networks achieved the best performance among four evaluated architectures.The evaluated architectures included GoogLeNet, AlexNet, VGG16, and a face-oriented deep network.
- 3.3. Patch-based Classification Stage: The reported experiments used 40× magnification, which produced the best performance among the evaluated magnification levels.
- 3.3. Patch-based Classification Stage: Hard-negative enrichment added tumor-adjacent or histologic-mimic regions that initially generated false positives and retrained the model.
- 3.4. Slide-based Classification: Slide classification aggregates 28 heatmap-derived geometric and morphological features into a whole-slide tumor probability.
- 3.4.2. Lesion-based Detection: Lesion localization thresholds the D-I heatmap, finds connected components, and averages D-I and D-II predictions over each component.D-II uses tumor-adjacent negatives to reduce false positives while reducing sensitivity.
4. Experimental Results
The system won both Camelyon16 evaluation tasks, achieving strong slide-level classification and lesion-level detection performance. Its slide-level predictions also complemented human diagnoses, substantially improving combined classification accuracy.
- Slide-based Evaluation: 0.9250 AUC was achieved for slide-based classification in Camelyon16.The evaluation measured discrimination between slides containing metastasis and normal slides using ROC AUC.
- Lesion-based Evaluation: The lesion-based evaluation used an FROC curve plotting sensitivity against average false positives per image.This task assessed detection of cancer lesions within whole slide images.
- Lesion-based Evaluation: When average false positives exceeded 2 per image, the system’s sensitivity outperformed the pathologist’s.The pathologist made no false-positive predictions, whereas the system surpassed the pathologist at higher false-positive rates.
- Human Comparison: 0.9948 AUC was obtained by combining the deep learning system’s predictions with the human pathologist’s diagnoses.The pathologist alone achieved an AUC of 0.9664; the combined result reduced the error rate from 3.4 percent to 0.52 percent.
5. Discussion
The discussion presents a deep learning system for metastatic cancer detection that combines deep network modeling with hard-negative enrichment and task-specific post-processing. The system achieved near-human classification performance, while combining its predictions with a pathologist substantially reduced diagnostic error.
- Method: Hard-negative enrichment targeted normal lymph-node regions initially misclassified as cancer.The system also used a state-of-the-art architecture and task-specific post-processing for slide classification and lesion detection.
- Related Work: Earlier digital pathology methods typically combined low-level image analysis with classical machine-learning classifiers using relatively small feature sets.The discussion contrasts these historical approaches with deep learning-based analysis.
- Human Comparison: Deep learning errors were not strongly correlated with human pathologist errors.This enabled the combined system to reduce the pathologist’s error rate from over 3 percent to less than 1 percent.