Workshop: Abstracts - Grand Challenge

MICCAI workshop - team abstracts

14.55: Team Aksell (speaker: Shujun He; 19th place QWK 0.9274)

Prostate cancer grade assessment of whole-slide images using, tile segmentation, self-attention, and multi-tasking learning

Developing accurate models to automatically diagnose prostate cancer is of crucial importance as prostate cancer is the second most common cancer among males worldwide. Here we describe a deep learning approach for automatic prostate cancer diagnosis that incorporates tile segmentation on low-resolution images, self-attention, and multi-tasking learning. Our methods are conceptually simple but effective, leading to accurate diagnosis of prostate cancer (private test quadratic weighted kappa = 0.927), despite noisy and imbalanced training data. In addition to accurate diagnosis, our multi-tasking deep learning model is interpretable due to usage of self-attention, and will only improve as more multi-labeled data accumulates.

15.05: Team Dmitry A. Grechka (speaker: Dmitry Grechka; 17th place QWK 0.9283)
How to predict the ISUP grade of prostate biopsies scans with a combination of CNN and RNN
Presenting the approach to building an automated system for grading the microscopy scans of biopsied tissue. The system consists of convolutional neural network (CNN) and a recurrent neural network (RNN) chained together to predict the ISUP grade group of a tissue sample. High resolution microscopy scan is split into a grid of smaller square tiles. Each tile containing the tissue is mapped into a feature vector by applying the CNN (DenseNet121). Then feature vectors (presented as a sequence) are passed to RNN (GRU units) to evaluate the presence of cancerous tissue and to assign a corresponding grade group.

15.15: Team UCLA Computational Diagnostics Lab (speaker: Wenyuan Li; QWK 0,9286)
Gleason grading of biopsies using an attention-based multi-resolution model ensembled with LGBM and XGBoost
We developed an automated prostate Gleason grading algorithm based on an attention-based multi-resolution model ensembled with lgbm and xgboost. Our model is trained on patch-based tissue samples extracted from whole slide images (WSI). A two-stage attention-based multiple instance learning (MIL) model using weakly supervised region of interest (ROI) detection was developed for ISUP-grade prediction. It is trained on multiple resolutions, with the lower resolution to identify suspicious regions that are further examined at higher resolution. To make the model more robust, we ensemble the MIL model with LGBM and xgboost models, whose feature extractors are trained to predict the primary and secondary Gleason scores.

15.25: Team lafoss (speaker: Maxim Shugaev; 11th place QWK 0.9301)
Concatenate tile pooling approach for end-to-end Gleason grading of biopsies
End-to-end system for automatic assessment of ISUP grade based on biopsy images is built. A novel Concatenate Tile pooling approach has been developed and allowed us to perform training based on labels assigned to entire images while using a tile-based approach. The computational cost of training is reduced by more than 10 times in comparison with training on full images. The training efficiency is further improved by use of a newly proposed tile cutout method. Progressive label self-distillation and removal of noisy labels is applied to handle the training data with a substantial level of label noise (0.853 QWK).

16.10: Team ChienYiChi (speaker: Jianyi Ji; 8th place QWK 0.9323)
Gleason grading of biopsies using attention and NetVLAD
Whole Slide Images (WSI) of biopsies have billions of pixels.The common way to deal with this kind of image is to extract some good tiles from it. To efficiently select potential tiles from WSI, we train a model with an attention layer over these tile candidates. Top k of the tiles are processed by another CNN model. Finally, all features aggregate with a NetVLAD layer before the final output.

16.20: Team BarelyBears (Speaker: Hiroshi Yoshihara; 6th place QWK 0.9326)
A label noise robust ISUP grading system using multi-instance learning.
We developed an automated ISUP grading system, which is an ensemble model of four multi-instance learning (MIL) networks. A MIL network consists of a feature extractor which extracts features from patches obtained from a whole slide image, and a head which concatenates and pools all the features and predicts the ISUP grade. Various backbones were used in the feature extractors. Networks were trained with Online Uncertainty Sample Mining, or with Mixup in order to improve robustness to label noises. The ensemble model trained noise-robustly showed better performance compared to the model trained ordinarily.

16.30: Team Save The Prostate (speaker: Rui Guo; 2nd place QWK 0.9377)
Gleason grading of biopsies with tile-based inputs
Deep learning methods have shown promising results in diagnosing prostate cancer (PCa). However, due to the various input shapes and giga-pixel resolution of patient biopsies, traditional training methods require extensive and costly computational resources. Our team has utilized a tile-based method to address these issues, which requires significantly low computational resources while maintaining a state of the art performance. Our approach includes a combination of 3 different methods: (1) Multi-Instance Learning (MIL) based CNN with attention-selected high-resolution input. (2) MIL based CNN with Squeeze-and-Excite Module across all tiles. (3) Deep CNN with re-composed single rectangular images from tiles. Additionally, we successfully tackled high-level label noise in training data by utilizing robust loss functions and pseudo labels.

16.40: Team PND (speaker: Yusuke Fujimoto; 1st place QWK 0.9409)
Gleason grading of biopsies with simple label noise reduction technique
We propose a simple label noise reduction technique, where we predict the Gleason score based on hold-out training, and remove noisy label data with a large gap between prediction and the ground truth.
While our method is very simple, yet useful; it can be easily applied to other tasks and models. We were able to win the competition by applying this method to simple EfficientNet based networks.