|
--- |
|
license: mit |
|
--- |
|
|
|
# Revisiting M3D is all you need |
|
[**Reference Paper**](https://arxiv.org/abs/2404.00578) | [**Code**](https://github.com/babbu3682/AMOS-MM-MI2RL) |
|
|
|
# Solution Overview |
|
This code is the 3rd place solution for the MICCAI 2024 AMOS-MM Challenge. It is based on the M3D model and utilizes the LLaMA-3.1-Instruct-7B. |
|
|
|
# Novelty |
|
* We divided the body into three regions: chest, abdomen, and pelvis. This allowed us to focus more effectively on these areas and further improve performance through text prompt engineering. |
|
* The introduction of LLaMA-3.1-8B-Instruct significantly boosted MRG and VQA performance. |
|
* Additionally, the use of an error note strategy led to further improvements in the final results. |
|
|
|
# Acknowledgement: |
|
We appreciate open source projects including: M3D, LLaVA, LLaMA. |
|
|
