Skip to content

Latest commit

 

History

History
68 lines (46 loc) · 3.24 KB

patch_classification.md

File metadata and controls

68 lines (46 loc) · 3.24 KB

Paper Information

Paper Title: Weakly supervised multiple instance learning histopathological tumor segmentation

Conference: MICCAI 2020

Official Code: https://github.com/marvinler/tcga_segmentation

Network Structure

Image

Short Summary

Within this paper a patch-based segmentation algorithm is proposed based on binary image-level labels of WSIs. MIL is applied to train an instance classifier model. Therefore, a WSI is divided into equally-sized patches (224 x 224) which are interpreted as instances.

Three questions about this paper:

1. [Problem Definition] What problem is this paper trying to solve?

It tackles the problem of WSI segmentation using only binary image-level labels of WSI patches

2. [Method] How does this paper solve the above problems?

It trains a binary patch classifier (ResNet50) which is then used to perform a segmentation of the gigapixel WSI. To train this model, a new method is proposed which enables to utilize patches with uncertain labels. This includes all patches from WSIs which are labelled to have cancerous tissue, since the extracted patches could also contain no cancerous tissue. During training, a single forward pass batch contains only a subsample of 150 images from a WSI. Label 1 (= cancerous) is assigned to 20% (= alpha) of the patches for which the network yields the highest probability. Label 0 (= non-cancerous) is assigned to 20% (= beta) of the patches for which for which the network yields the lowest probability. Only these patches are then used to update the weights. The rest is discarded.

3. Details about the experiment

3.1 Which Datasets are used?

  • Image type: flash-frozen H&E stained WSIs, kidney, lung, and breast cancer (TCGA)
  • Image number: 6212
  • Train/Val/Test: 4212 / 972 / 1296
  • Image size: 100k x 100k, but 224 x 224 patches were sampled from the WSIs for training
  • Resolution: 0.5 microns/pixel at 20x magnification

To measure out-of-distribution performance or generalization ability, the model was evaluated on additional 100 slides extracted from locations that were not used during training (colon, ovary, corpus uteri). Additionally, the model was evaluated on the PatchCamelyon dataset, which contains formalin-fixed, parafin-embedded (FFPE) tissues. These are visually different from flash-frozen ones.

3.2 What are the implementation details?

  • Hardware: 2 NVIDIA V100 GPUs
  • Training Time: 20 epochs, 16 hours training
  • Architecture: ResNet50 with 15 different configurations for alpha and beta were trained
  • Data augmentation: Rotation, Mirroring, Color jitter, channel-wise standard scaling from training averages and variances
  • Initialization: Pretrained on ImageNet
  • Batch size: 150

3.3 What's the evaluation metric?

  • Precision
  • Recall
  • F1-Score
  • AUC

3.4 What is the ranking of the experiment results?

No comparision with other methods was performed. The following table shows the AUC values for different configurations. Grey results take background into account.

Image

Performance on PatchCamelyon is not included in the table above and achieved the optimal AUC of 0.802 for alpha = 20% and beta = 20%.