Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Efficient Quality Control of Whole Slide Pathology Images with Human-in-the-loop Training

Abhijeet Patil, Harsh Diwakar, Jay Sawant, Nikhil Cherian Kurian, Subhash Yadav, Swapnil Rane, Tripti Bameta, Amit Sethi

TL;DR

HistoROI is introduced, a robust yet lightweight deep learning-based classifier to segregate WSI into 6 broad tissue regions-epithelium, stroma, lymphocytes, adipose, artifacts, and miscellaneous, and consistently performs well across multiple organs, despite being trained on only a single dataset, demonstrating strong generalization.

Abstract

Histopathology whole slide images (WSIs) are being widely used to develop deep learning-based diagnostic solutions, especially for precision oncology. Most of these diagnostic softwares are vulnerable to biases and impurities in the training and test data which can lead to inaccurate diagnoses. For instance, WSIs contain multiple types of tissue regions, at least some of which might not be relevant to the diagnosis. We introduce HistoROI, a robust yet lightweight deep learning-based classifier to segregate WSI into six broad tissue regions -- epithelium, stroma, lymphocytes, adipose, artifacts, and miscellaneous. HistoROI is trained using a novel human-in-the-loop and active learning paradigm that ensures variations in training data for labeling-efficient generalization. HistoROI consistently performs well across multiple organs, despite being trained on only a single dataset, demonstrating strong generalization. Further, we have examined the utility of HistoROI in improving the performance of downstream deep learning-based tasks using the CAMELYON breast cancer lymph node and TCGA lung cancer datasets. For the former dataset, the area under the receiver operating characteristic curve (AUC) for metastasis versus normal tissue of a neural network trained using weakly supervised learning increased from 0.88 to 0.92 by filtering the data using HistoROI. Similarly, the AUC increased from 0.88 to 0.93 for the classification between adenocarcinoma and squamous cell carcinoma on the lung cancer dataset. We also found that the performance of the HistoROI improves upon HistoQC for artifact detection on a test dataset of 93 annotated WSIs. The limitations of the proposed model are analyzed, and potential extensions are also discussed.

Efficient Quality Control of Whole Slide Pathology Images with Human-in-the-loop Training

TL;DR

HistoROI is introduced, a robust yet lightweight deep learning-based classifier to segregate WSI into 6 broad tissue regions-epithelium, stroma, lymphocytes, adipose, artifacts, and miscellaneous, and consistently performs well across multiple organs, despite being trained on only a single dataset, demonstrating strong generalization.

Abstract

Histopathology whole slide images (WSIs) are being widely used to develop deep learning-based diagnostic solutions, especially for precision oncology. Most of these diagnostic softwares are vulnerable to biases and impurities in the training and test data which can lead to inaccurate diagnoses. For instance, WSIs contain multiple types of tissue regions, at least some of which might not be relevant to the diagnosis. We introduce HistoROI, a robust yet lightweight deep learning-based classifier to segregate WSI into six broad tissue regions -- epithelium, stroma, lymphocytes, adipose, artifacts, and miscellaneous. HistoROI is trained using a novel human-in-the-loop and active learning paradigm that ensures variations in training data for labeling-efficient generalization. HistoROI consistently performs well across multiple organs, despite being trained on only a single dataset, demonstrating strong generalization. Further, we have examined the utility of HistoROI in improving the performance of downstream deep learning-based tasks using the CAMELYON breast cancer lymph node and TCGA lung cancer datasets. For the former dataset, the area under the receiver operating characteristic curve (AUC) for metastasis versus normal tissue of a neural network trained using weakly supervised learning increased from 0.88 to 0.92 by filtering the data using HistoROI. Similarly, the AUC increased from 0.88 to 0.93 for the classification between adenocarcinoma and squamous cell carcinoma on the lung cancer dataset. We also found that the performance of the HistoROI improves upon HistoQC for artifact detection on a test dataset of 93 annotated WSIs. The limitations of the proposed model are analyzed, and potential extensions are also discussed.
Paper Structure (8 sections, 6 figures, 1 table)

This paper contains 8 sections, 6 figures, 1 table.

Table of Contents

  1. Introduction
  2. Datasets
  3. METHODOLOGY
  4. RESULTS AND DISCUSSION
  5. HistoROI as WSI pre-processing tool
  6. Validation of HistoROI on CRC-100k dataset
  7. HistoROI as QC tool
  8. Conclusion

Figures (6)

  • Figure 1: An example of segmentation results on a whole slide image using HistoROI into epithelium, stroma, lymphocytes, adipose, artifacts, and miscellaneous areas
  • Figure 2: Human-in-the-loop training pipeline for HistoROI: Actions in red boxes are automatic while those in green boxes are manual. (a) Embeddings of the patches of WSI are divided into clusters. (b) Clusters are manually annotated. Heterogeneous clusters are re-clustered (shown using dotted line) (c) Newly annotated data is added to previously annotated data and HistoROI is trained with the updated data. (d) The trained model is applied (tested) on multiple WSIs. WSIs with poor performance are manually identified (denoted by X) and annotated for the next iteration of training.
  • Figure 3: HistoROI quantitative results: (a) Box plot of AUCs on the test split of the CAMELYON16 dataset shows improvement in performance when WSIs are pre-processed with HistoROI. (b) Box plot of AUCs for the TCGA-LUNSC dataset. (c) Comparison of HistoROI with HistoQC histoqc for QC using TCGA-4Org dataset shows that HistoROI's Dice score is better than HistoQC for 65 WSIs out of 93 WSIs. (d) HistoROI predicts 77% of patches correctly in the CRC-100k without being trained on this dataset.
  • Figure 4: Qualitative results on CAMELYON16: CLAM clam trained on regions filtered by HistoROI generates better attention maps. Metastasised region is annotated with cyan color in the first column. CLAM attention map for the "All" strategy is scattered compared to the accurate and specific one generated using the "QCFat--" strategy.
  • Figure 5: Qualitative results on CRC-100kcrc100k: Example predictions (column labels) on CRC-100k dataset by HistoROI for various classes (row labels) shows that misclassified patches (with red crosses) tend to have background, staining issues, or multiple classes, which are not present in the correct classified patches (with green ticks). Here BACK stands for the background, ADI for adipose, MUC for mucus, NORM for normal, TUM for tumor, STR for stroma, MUS for muscle, DEB for debris, and LYM for lymphocytes.
  • ...and 1 more figures