Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Parameter-efficient Fine-tuning for improved Convolutional Baseline for Brain Tumor Segmentation in Sub-Saharan Africa Adult Glioma Dataset

Bijay Adhikari, Pratibha Kulung, Jakesh Bohaju, Laxmi Kanta Poudel, Confidence Raymond, Dong Zhang, Udunna C Anazodo, Bishesh Khanal, Mahesh Shakya

TL;DR

The paper tackles automated brain tumor segmentation in Sub-Saharan Africa by addressing domain shift between BraTS-2021 and BraTS Africa and data scarcity in LMICs. It introduces a Parameter-Efficient Fine-Tuning (PEFT) approach that plugs convolutional adapters into the MedNeXt-S backbone, allowing effective adaptation using BraTS-2021 for pretraining and BraTS Africa for fine-tuning. Results show PEFT achieves mean Dice near 0.80 on BraTS Africa, closely matching full fine-tuning (≈0.77) and outperforming training solely on BraTS Africa (≈0.72), while substantially reducing training time and parameter overhead. The method exhibits high specificity (≈0.99) but lower sensitivity (≈0.75), suggesting a tendency to over-segment in some regions, yet provides a practical, resource-efficient avenue for domain-adaptive brain tumor segmentation in low-resource clinical settings. The work includes a public codebase for replication and further exploration of adapter placements and qualitative outcomes.

Abstract

Automating brain tumor segmentation using deep learning methods is an ongoing challenge in medical imaging. Multiple lingering issues exist including domain-shift and applications in low-resource settings which brings a unique set of challenges including scarcity of data. As a step towards solving these specific problems, we propose Convolutional adapter-inspired Parameter-efficient Fine-tuning (PEFT) of MedNeXt architecture. To validate our idea, we show our method performs comparable to full fine-tuning with the added benefit of reduced training compute using BraTS-2021 as pre-training dataset and BraTS-Africa as the fine-tuning dataset. BraTS-Africa consists of a small dataset (60 train / 35 validation) from the Sub-Saharan African population with marked shift in the MRI quality compared to BraTS-2021 (1251 train samples). We first show that models trained on BraTS-2021 dataset do not generalize well to BraTS-Africa as shown by 20% reduction in mean dice on BraTS-Africa validation samples. Then, we show that PEFT can leverage both the BraTS-2021 and BraTS-Africa dataset to obtain mean dice of 0.8 compared to 0.72 when trained only on BraTS-Africa. Finally, We show that PEFT (0.80 mean dice) results in comparable performance to full fine-tuning (0.77 mean dice) which may show PEFT to be better on average but the boxplots show that full finetuning results is much lesser variance in performance. Nevertheless, on disaggregation of the dice metrics, we find that the model has tendency to oversegment as shown by high specificity (0.99) compared to relatively low sensitivity(0.75). The source code is available at https://github.com/CAMERA-MRI/SPARK2024/tree/main/PEFT_MedNeXt

Parameter-efficient Fine-tuning for improved Convolutional Baseline for Brain Tumor Segmentation in Sub-Saharan Africa Adult Glioma Dataset

TL;DR

The paper tackles automated brain tumor segmentation in Sub-Saharan Africa by addressing domain shift between BraTS-2021 and BraTS Africa and data scarcity in LMICs. It introduces a Parameter-Efficient Fine-Tuning (PEFT) approach that plugs convolutional adapters into the MedNeXt-S backbone, allowing effective adaptation using BraTS-2021 for pretraining and BraTS Africa for fine-tuning. Results show PEFT achieves mean Dice near 0.80 on BraTS Africa, closely matching full fine-tuning (≈0.77) and outperforming training solely on BraTS Africa (≈0.72), while substantially reducing training time and parameter overhead. The method exhibits high specificity (≈0.99) but lower sensitivity (≈0.75), suggesting a tendency to over-segment in some regions, yet provides a practical, resource-efficient avenue for domain-adaptive brain tumor segmentation in low-resource clinical settings. The work includes a public codebase for replication and further exploration of adapter placements and qualitative outcomes.

Abstract

Automating brain tumor segmentation using deep learning methods is an ongoing challenge in medical imaging. Multiple lingering issues exist including domain-shift and applications in low-resource settings which brings a unique set of challenges including scarcity of data. As a step towards solving these specific problems, we propose Convolutional adapter-inspired Parameter-efficient Fine-tuning (PEFT) of MedNeXt architecture. To validate our idea, we show our method performs comparable to full fine-tuning with the added benefit of reduced training compute using BraTS-2021 as pre-training dataset and BraTS-Africa as the fine-tuning dataset. BraTS-Africa consists of a small dataset (60 train / 35 validation) from the Sub-Saharan African population with marked shift in the MRI quality compared to BraTS-2021 (1251 train samples). We first show that models trained on BraTS-2021 dataset do not generalize well to BraTS-Africa as shown by 20% reduction in mean dice on BraTS-Africa validation samples. Then, we show that PEFT can leverage both the BraTS-2021 and BraTS-Africa dataset to obtain mean dice of 0.8 compared to 0.72 when trained only on BraTS-Africa. Finally, We show that PEFT (0.80 mean dice) results in comparable performance to full fine-tuning (0.77 mean dice) which may show PEFT to be better on average but the boxplots show that full finetuning results is much lesser variance in performance. Nevertheless, on disaggregation of the dice metrics, we find that the model has tendency to oversegment as shown by high specificity (0.99) compared to relatively low sensitivity(0.75). The source code is available at https://github.com/CAMERA-MRI/SPARK2024/tree/main/PEFT_MedNeXt

Paper Structure

This paper contains 16 sections, 4 equations, 7 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Transfer Learning and Fine-tuning
  4. Parameter Efficient Fine-tuning
  5. Methodology
  6. Overview
  7. Dataset
  8. Preprocessing
  9. Model Architecture
  10. Proposed Architecture with Adapter
  11. Results and Discussion
  12. Conclusion
  13. Acknowledgement
  14. Supplementary Materials
  15. Performance of Parallel Vs Sequential ConvNeXt Adapter
  16. ...and 1 more sections

Figures (7)

  • Figure 1: Brain image slices of a representative case from the BraTS Africa dataset with the four MRI modalities and manual annotated subregions (Mask), representing brain tumor sub-regions: Left to Right; T1-contrast-enhanced (T1c), pre-contrast T1-weighted (T1w), FLAIR, T2-weighted (T2w), and Mask
  • Figure 2: Left: ConvNeXt-Adapter. Right: Different forms of Adapter placement within the MedNeXT Superblock
  • Figure 3: Conv-Adapter Inspired MedNext Segmentation Model Adapter Architecture
  • Figure 4: Boxplot comparison of segmentation methods: without fine-tuning, full fine-tuning, and PEFT, using Average Dice (left) and Average HD95 (right)
  • Figure 5: Visual comparison of Predicted and Ground Segmentations with and without finetuning.
  • ...and 2 more figures