tyrellto/ATD-challenge

Attempt at the RSNA 2023 ATD Challenge on Kaggle

ATD_challenge

---

Table of Contents:

• Introduction
• Methods Employed
  • 3D Segmentation with U-Net + ResNet18
  • Bounding Box Extraction
  • Organ Detection
  • Organ Extraction
  • 2.5D LSTM + CNN (ConvNext-Nano) Training
• Future Work and Considerations

Introduction

This repository contains my attempt at a challenging medical image processing task for RSNA 2023 Abdominal Trauma Detection. The goal was to effectively utilize a combination of 3D segmentation and 2.5D LSTM + CNN models to achieve reliable results. Throughout the project, various methods and strategies were employed to preprocess, segment, and classify medical images.

Methods Employed

1. 3D Segmentation with U-Net + ResNet18:

   • The initial stage involved preprocessing the given segmentation masks and volume data.
   • A 3D segmentation model using U-Net with a ResNet18 backbone was trained. This process is resource-intensive and requires significant GPU VRAM.
   • This trained model was then used to segment other series/patients in the dataset.

2. Bounding Box Extraction:

   • Post segmentation, bounding boxes were extracted from the segmentation masks to identify relevant slices.
   • Non-overlapping images were created using 3 adjacent slices. For instance, image_0 is derived from slices 0, 1, and 2, while image_1 is derived from slices 3, 4, and 5.

3. Organ Detection:

   • To determine missing organs in some series, a mapping was generated showing which organs were present in specific slices.
   • Linear regression models were trained for each organ to predict the presence of an organ in a given slice.
   • Combinations used for training:

combinations = [
    (["liver", "bowel"], "kidneys"),
    (["liver", "spleen", "bowel"], "kidneys"),
    ...
    (["liver"], "spleen")
]

4. Organ Extraction:

   • Based on the predictions and the slices where organs were detected, the liver, kidneys, spleen, and bowel were extracted from their respective subregions.
   • These extracted subregions were preprocessed and resized to a uniform size of 10 x 320 x 320, primarily due to GPU VRAM limitations.

5. 2.5D LSTM + CNN (ConvNext-Nano) Training:

   • The final model architecture utilized was a combination of LSTM and ConvNext-Nano, trained on the processed organ subregions. The model worked with an image size of 320x230.
   • Given the class imbalance in the dataset, there was an effort to weigh the underrepresented classes more. However, this aspect requires further fine-tuning.

Future Work and Considerations

1. Improved 3D Segmentation: One key area of improvement is to further train and refine the 3D segmentation model to achieve better segmentation masks, especially for specific organs.
2. Slice Selection Methodology: A more sophisticated slice selection method can potentially yield better results.
3. Utilizing Segmentation Masks: In the training phase of the 2.5D LSTM + CNN model, incorporating the segmentation masks directly might offer improved performance.

---