Kaggle RSNA Breast Cancer Detection - Silver Medal Solution

This project focuses on detecting breast cancer based on CT imaging data, formulated as a binary classification problem. The competition evaluates solutions using the pF1 score.

Steps for the Solution

1. Data Extraction

The CT images cannot be directly used in standard formats, so we extract them using the NVIDIA DALI framework for fast processing.

• Run: step1_extract_raw_data.ipynb
• This step mirrors the code from the open-source solution.

2. Cropping the Images

To remove unnecessary blank areas from the CT scans, edge detection is employed to identify and crop out irrelevant parts, focusing on the breast region.

• Run: step2_crop_data.ipynb

3. Convert to TFRecord

Transform the cropped image data into TFRecord format for efficient loading during model training.

• Run: step3_tfrecord.ipynb

4. Model Training

We train the model using ConvNextV2, employing data augmentation techniques (rotation, brightness adjustments, flipping, and cropping). The training is optimized using the AdamW optimizer with Stochastic Weight Averaging (SWA) across five cross-validation folds.

• Run: step4_train.ipynb

5. Model Inference

Using the five trained models, we generate predictions. The results are averaged, and the top 2.1% quantile of probability values is used as the threshold for binary classification.

• Run: step5_infer.ipynb

Final Result

The ensemble model achieves a final ranking of top xx%, placing in the top xxx4% of the competition.