Pap Smear Cell Classification Challenge (PS3C)

This is the repository for the paper "Towards Improved Cervical Cancer Screening: Vision Transformer-Based Classification and Interpretability" at PS3C challenge of team Khoa_NT.

---

Data

Download the data follow this instructions and put it in the data/ folder. The data should be organized as follows:

dataset
│
├── isbi2025-ps3c-test-dataset
│   ├── isbi2025_ps3c_test_image_00001.png
│   ├── isbi2025_ps3c_test_image_00002.png
│   ├── ...
│   ├── isbi2025_ps3c_test_image_18158.png
│   └── isbi2025_ps3c_test_image_18159.png
│
├── isbi2025-ps3c-train-dataset
│   ├── bothcells
│   ├── healthy
│   ├── rubbish
│   └── unhealthy
│
└── pap-smear-cell-classification-challenge
    ├── isbi2025-ps3c-test-dataset.csv
    └── isbi2025-ps3c-train-dataset.csv

Training

Multi-class classification

./train.sh

Will train the model and save the best model to ckpt/ folder.

Binary classification

./train_binary.sh

Will train the model and save the best model to ckpt/ folder. Check train_binary.sh for example usage.

Features Classification

Using the extracted features to train the classifier.

Training

./train_features_classifier.sh

Inference

./infer_features_classifier.sh

There are two types of outputs:

• predictions_<csv_path>.csv: Predictions with probabilities for each class, saved to the folder containing the ckpt file.
• submission_<ckpt_path>.csv: Predictions for submission, saved to submission/feature_classifier/ folder.

Submission

Multi-class classification

./predict.sh

Will predict on the test set and save the results for submission to submission/ folder. Also, will save the predicted results to the folder containing the ckpt file. Check predict.sh for example usage.

Binary classification (☀️/🌙)

./predict_binary.sh

Will predict on the test set and save the results for submission to submission/ folder. Also, will save the predicted results to the folder containing the ckpt file. Check predict_binary.sh for example usage.

Note: After predicting all binary models, run merge_binary_predictions.py to merge the results in submission/ folder. For example, If we train the 1st model with rubbish vs healthy_unhealthy_bothcells, the predicted healthy_unhealthy_bothcells of the 1st model will be replaced by the prediction (healthy vs unhealthy_bothcells) of the 2nd model.

• Check merge_predictions.sh & merge_binary_predictions.py for example usage.
• Check the Binary Classification Result below to see the merged result.

Analysis

Predicting on training set

./predict_train.sh

Will predict on the training set and save the results to the folder containing the ckpt file. If no ckpt file is provided, will use the pretrained model and save the results to the folder with the name of the pretrained model.

Extracting features from the model

./extract_features.sh

Will extract features from the model and save the results to extracted_features/{model_name} folder. If no ckpt file is provided, will use the pretrained model and save the results to the folder with the name of the pretrained model.

Extract modes:

• pooled: Extract features from the pooled output of the model. Usually pooled without the classifier token.
• pooled_all: Extract features from all outputs of the model. Including the classifier token.
• classifier_token: Extract features from the classifier token.

Extract features Google Drive link (Temporary):

• eva02_base_patch14_448: Features from the model overfit on the training set.
• PreTrained_eva02_base_patch14_448: Features from the ImageNet pretrained model.

Binary Classification Result

Name	Model 1	Model 2	Model 3
rubbish	☀️	-	☀️
healthy	🌙	☀️	🌙
unhealthy	🌙	🌙	-
bothcells	🌙	🌙	-

Model Combination	Test Result	Comment
1 + 2	0.84982
3 + 2	0.85025
3	0.84127	🤔 Get 0.84 with only submitting rubbish & healthy (without unhealthy & bothcells)

Importance-based Feature Selection

./importance_based_feature_selection/mask.sh

Will create a masking table based on the feature importances of machine learning models, including Logistic Regression, Gradient Boosting, and Random Forest. This masking table will then be used to perform importance-based feature selection.

TODO

• Extracting features from the model
• Add batch size for prediction script
• Use WandB to log training and prediction
• Binary classification strategy: Rubbish vs (Healthy/Unhealthy) -> Healthy vs Unhealthy
• Training with the training set based on csv files (Idea 1)
• Add class weights to the loss function (Idea 2)
• Add SAE strategy: training and extracting SAE features from the extracted features (Classifier token, Pooled, Pooled_all)
• Add training and predicting with features (Idea 3)
• Add Validation set / Cross-validation
• Docker

Ideas

• Idea 1: Remove confusing data by identifying the file names of False Positives (FP) and False Negatives (FN) and excluding them from the training set.
• Idea 2: Since classifying unhealthy cases is key to success, assign a higher loss weight to misclassifications of unhealthy cases.
• Idea 3: Create the mask to prune the extracted features (e.g., Eva02 or SAE's features). Then train the classifier with the pruned features.

Docker

Khoa's environment is available in docker/khoa/khoa.md.