Multi-Teacher Cross-Modal Distillation with Cooperative Deep Supervision Fusion Learning for Unimodal Segmentation

Abstract

Accurate brain tumor segmentation is a labor-intensive and time-consuming task that requires automation to enhance its efficacy. Recent advanced techniques have shown promising results in segmenting brain tumors; however, their dependency on extensive multimodal magnetic resonance imaging (MRI) data limits their practicality in clinical environments where such data may not be readily available. To address this, we propose a novel multi-teacher cross-modal knowledge distillation framework, which utilizes the privileged multimodal data during training while relying solely on unimodal data for inference. Our framework is tailored to the unimodal segmentation of the $T_{1ce}$ MRI sequence, which is prevalently available in clinical practice and structurally akin to the $T_{1}$ modality, providing ample information for the segmentation task. Our framework introduces two learning strategies for knowledge distillation (KD): (1) performance-aware response-based KD and (2) cooperative deep supervision fusion learning (CDSFL). The first strategy involves dynamically assigning confidence weights to each teacher model based on its performance, ensuring that the KD is performance-driven, and the CDSFL module augments the learning capabilities of the multi-teacher models by fostering mutual learning. Moreover, the fused information is distilled into the student model to improve its deep supervision. Extensive experiments on BraTS datasets demonstrate that our framework achieves promising unimodal segmentation results on the $T_{1ce}$ and $T_{1}$ modalities and outperforms previous state-of-the-art methods.

---

Contributions

The contribution of our work is stated below:

• Proposed a novel CDSFL module for cooperative deep supervision learning among teacher models that incorporates attention mechanisms to contextualize the feature maps and enable the teacher models to learn from each other. The fused knowledge is then distilled to the student model to improve its representation.
• Introduced customized performance-aware response-based KD scheme for the multi-teacher framework that dynamically allocates weights to each teacher model to distill its knowledge based on its performance.
• Proposed a novel MTCM-KD framework for unimodal brain tumor segmentation. The framework combines the concepts of performance-aware response-based KD and CDSFL to increase its learning capability.
• Performed extensive experiments and evaluated the framework on the BraTS datasets from 2018 to 2021. Our approach yields promising results and surpasses previous state-of-the-art models in unimodal brain tumor segmentation of $T_{1}$ and $T_{1ce}$ modalities.

---

Model Architecture

Schematic diagram of the proposed MTCM-KD framework.

---

Dataset

This work uses four BraTS datasets from 2018-2021. The first three BraTS datasets can be downloaded from here, While the BraTS-2021 dataset is available here.

---

Setup

Install the requirements

pip install requriements.txt

and if in case there is still an issue installing the cudatoolkit and GPU version, then refer to the site Pytorch or you can use the below commands for installing the Pytorch along with cudatoolkit

• Linux / Window

  pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118

• Mac

  pip3 install torch torchvision torchaudio

Update and customize the model's configuration by making the required changes in the config.json file and running the code using the below command

python run.py

For the inference of the model, make the necessary changes in the inference.py file, and then run the script:

python inference.py

---

3D Visualization

To visualize the 3D MRI volumes of the brain tumor, first, predict the 3D volumes of MRI by executing the below script:

python visualize.py

• then refer to the repository for the 3D visualization of the volumes: 3D nii Visulalizer

Citation

@article{AHMAD2024111854,
title = {Multi-teacher cross-modal distillation with cooperative deep supervision fusion learning for unimodal segmentation},
journal = {Knowledge-Based Systems},
volume = {297},
pages = {111854},
year = {2024},
issn = {0950-7051},
doi = {https://doi.org/10.1016/j.knosys.2024.111854},
url = {https://www.sciencedirect.com/science/article/pii/S095070512400488X},
author = {Saeed Ahmad and Zahid Ullah and Jeonghwan Gwak},
}