This repository presents the 1st place solution for the Liver Lesions Detection based on Ultrasound Image Hackathon held from August 12th to 14th, 2023, at Mahidol University.

The solution leverages the DINO architecture with a Swin Transformer backbone to detect three classes of liver lesions (cystic, fibrosis, solid) as well as normal liver images. Our solution achieved a mAP70 score of 0.5037 on the public leaderboard (using 25% of the test data) and 0.5263 on the private leaderboard (utilizing the entire test data).

• The Transformer backbone outperforms the convolutional backbone. We attribute this to the inherent noise present in ultrasound images. The Transformer's backbone is more adept at learning improved ultrasound image representations.
• Incorporating all images during training is crucial. In the mmdetection dataloader configuration, the option filter_cfg=dict(filter_empty_gt=False) is utilized. This allows the model to see more negative examples during training.

To train our solution using the provided dino-5scale_swin-l_8xb2-12e_liver.py config file, follow these steps:

1. Follow the instructions to install mmcv from the official documentation: mmcv Installation Guide
2. Clone the mmdetection repository:

   git clone https://github.com/open-mmlab/mmdetection.git
   

3. Download the COCO pretrained weights for SWIN-L from the provided link: SWIN-L Pretrained Weights
4. Modify the paths in the provided config file dino-5scale_swin-l_8xb2-12e_liver.py to correspond to your dataset and pretrained weights location.
5. Navigate to the mmdetection directory:

   cd mmdetection
   

6. Begin training with:

   python tools/train.py path/to/dino-5scale_swin-l_8xb2-12e_liver.py
   

For a detailed training configuration, refer to the file dino-5scale_swin-l_8xb2-12e_liver.py. Additional information on configuring models can be found in the mmdetection documentation.

• Zaw Htet Aung
• Kittinan Srithaworn
• Titipat Achakulvisut