# [8th Place Solution] Understanding Data Before Designing Methods

Acknowledgments
I would like to express my gratitude to Kaggle, the organizers, and other participants in the community. I have learned a great deal from this competition and hope that it will promote the advancement of research in MIL，Classification and outliers detection of Human tissues, and the study of women's health.
Introduction
The first and most crucial step in this competition is to familiarize oneself with the feature of Whole Slide Images (WSI) and Tissue Microarray (TMA) images. A simple visual examination reveals that
the features of TMA and WSI are presented at vastly different scales.
TMA features are at the cellular level, while WSI features are several to tens of times larger than cell clusters. Hence, from the onset of the competition, I decided to tackle WSI and TMA with two completely distinct approaches.
Key points for the global approach:
Address WSI and TMA separately
.
Align the magnification of TMA and WSI
to enable the reuse of TMA training code for optimizing the WSI feature extractor.
WSI images are 20x magnified, while TMA images are 40x. Thus, TMA images need to be downsampled by a factor of 2 to align with the physical scale of WSI, which can also be observed with a visual inspection of the training data.
Architecture of the solution
![](assets/c7d6b413a704100d82223395.png)
TMA Approach
Summary of Key Points:
Train with patches tiled with official masks
(most important globally).
Train the model with
healthy and death patches
to predict some outliers.
Employ
arcface
to retrieve some outliers and part of the 5-class classification.
Heavy ensemble of 6 classification models to predict samples that arcface retrieval did not cover.
Inference:
Use arcface retrieval with the training set's TMA as templates. Output 5-class results for samples with a close cosine distance, and consider those with a far cosine distance as outliers. Samples that are uncertain in the arcface stage are left for the subsequent phase.
Models: 5-fold effv2s + 5-fold convnext small with dynamic margin, subcenter=3.
Top1 threshold at 0.05, Other threshold at 0.2.
The 6-class model uses 2-fold effv2s + 5-fold effv2l + 4-fold convnext small + 3-fold convnext large.
Model
Resolution
Effv2s
1280
Effv2l
1280
Convnext Small
1024
Convnext Large
1024
Note: The arcface can tackle about 60% of the TMAs. So even the second stage is heavy, it doesn't cause timeout.
Training:
Tile patches using the official segmentation masks. In addition to the official 5 categories, classify patches of
healthy
and
dead
as
Other
class. Instead of training with the official TMAs , which has limited number, use them for subsequent validation and retrieval.
Use the first-phase models to generate pseudo-label patches on the remaining 300+ WSIs without mask.
Inherit the weights from the first or second step, training only the last layer of the backbone and the arcface head.
WSI Approach
Summary of Key Points:
Train the feature extractor using the TMA pipeline
.
Synthesize
Other
WSI
during the training process.
Integrate different magnification scales.
Rank the DataFrame of WSI with the number of pixels, and process the WSIs with multithread can greatly speed the the WSI processing  procedure.
Inference:
Ensemble of two resolutions of feature extractor: 3072 resized to 768, and 1024 without down-sampling. After extracting features, apply DTFD-MIL.
For speed up the image processing procedure, I only use the center region of each 3072 tile, so that I only need to crop patch from the WSI once.
Model
Resolution
Number folds
Convnext Small
3072 resize 768
2
Effv2s
3072 resize 768
3
Effv2s
1024
4
Training:
Train the feature extractor using the TMA pipeline.
Extract features using the feature extractor.
Predict the probability of 'Other' on all patches using the feature extractor, and
create an
Other pool
with patches that have a high probability of 'Other'.
During the training of DTFD-MIL, dynamically synthesize some
WSI
from the
Other pool
each epoch.
Training DTFD-MIL serves as a validation for whether the TMA pipeline models truly learned useful features. If  we use the pretrained weights from ImageNet, DTFD-MIL requires up to 200 epochs to converge. In contrast, using models trained on TMA, the MIL Head may take as little as 1 epoch and at most 20 epochs to converge.
Something I don't have time to try but I think may work
Arcface for WSI.
Large transformer pretrained on Large set of Slides. In fact I tried PLIP at the early stage of this competition,  But I didn't dig deeper.
Better retrieve strategy for arcface.
More resolution for WSI . I tried to add
6144 resize to 1024
into my final pipeline. But the notebook crashed.
Ensemble of MIL head
External data.
Todo:
This post is  almost completed :
If possible, I may add more ablation study in two weeks.
Discussion and Citation
[DTFD-MIL](https://arxiv.org/abs/2203.12081)
is a strong and robust baseline for MIL, which follows ABMIL.
I think MIL methods are not sensitive to the position and the number of patches, which is observed in my experiment. That's why I randomly synthesized some
WSI
from the
Other pool
each epoch for WSI. And reduce the number on
1024 resolution wo down-sample
for submission.  This is also proved in
[this paper](https://ieeexplore.ieee.org/document/10219719/authors#authors)
.
The reason why I ensemble multiple resolution for WSI: if you check the
[clarification on WHO](https://www.pathologyoutlines.com/topic/ovarytumorwhoclassif.html)
. You will find that Pathologists distinguish the subtypes of Ovarian cancer on different magnification. So I believe ensemble multiple resolution for classifying a WSI should be very important.
About Larger transformer-based model: I believe the competition needs model with better generalization. And larger models are often more robust. At the early stage of this competition, I takes some time to test PLIP offline, but the CV is not good. Also, since I decide to solo this time, So I have to spend my limited time on the direction I am most confident about, solving this task in a more traditional and solid way.
Closing Thoughts: My Journey to Grandmaster
The path to becoming a Grandmaster has been lengthy, filled with challenges, and an experience I will cherish for life. I am incredibly thankful for the support from my teammates in past competitions, as well as the unwavering encouragement from my family and girlfriend throughout my journey. The final step to becoming a Grandmaster, achieving a solo gold medal, has been particularly solitary and tough. This was my fourth attempt at a solo gold medal. if I failed this time, it might have been three years, ten years, or perhaps never before I'd have the chance again, as I am about to graduate with my master's degree and start my career in a busy company. Fortunately, I have realized the dream I had three years ago and have now brought my student years to a close with the title of Grandmaster. Wishing everyone a Happy New Year!