Multiple Instance Learning with Center Embedding for Histopathology Classification
Year: 2020, Author: Philip Chikontwe, Meejeong Kim, Sang Hyun Park et al.
Journal Name: MICCAI
1. Background
- Histopathology images play a vital role as a gold standard for cancer diagnosis in clinical settings.
- Pixel-level annotations are difficult, time-consuming and costly to acquire.
2. Problems & Solutions
-
Problems
- Instance labels in MIL is ambiguity and learning robust instance embeddings is also challenging. Some approach often fails when learning based on noisy labels.
-
Solutions
- Use an end-to-end framework that exploits instance- & bag-level representations.
3. Methods
- Overview of the proposed framework
- Notes
4. Experiments
Dataset
- Two colectoral cancer (CRC) WSI datasets involve H&E stained.
- Dataset Ⅰ consists of 173 WSIs (59 normal & 114 MSI).
- Dataset Ⅱ consists of 193 WSIs (85 normal & 108 MSI).
- 40% training, 10% validation, 50% testing.
Evalution Metrics
- Use Ostu method.
- https://blog.csdn.net/qq_29462849/article/details/81022607
Settings
- Set top-k instance selection to 50 and used instances with 256 X 256 size.
- Use finetuned ResNet-34 as the feature extraction model Fθ (·).
- Fully Conected (FC) layers as the instance and bag classifiers HI and HB.
- The feature dimension of the embeddings and center model was set to 512.
- Learning rate = 1e-4, #epoch = 40.
- Loss hyperparameters λ = 0.01, α = 1.0, β = 0.01.
5. Results
- Comparision Methods
- Qualitative Results
- Segmentation Results
6. Contributions
- Propose an end-to-end framework for wekly superwised histopathology classification to assign correct instance and bag labels.
- Show that center loss can improve bag features learning and also reduce ambiguity in instance labels.
- Propose a soft-assignment based inference method for bag-level prediction with overal improved performance over recent SOTA methods.