1 of 10

Lightweight Grey Wolf Optimization-Driven

Hyperparameter Tuning of U-Net

for Robust Brain Tumor Segmentation

Shoffan Saifullah | Rafał Dreżewski | Anton Yudhana

26th International Conference on Computational Science

29 June - 1 July 2026 • DESY • Hamburg • Germany

ICCS 2026

2 of 10

Why this problem matters

U-Net works well, but its performance depends heavily on tuning.

Brain tumor MRI segmentation requires accurate masks, yet manual annotation is time-consuming and variable.

U-Net is effective, but depth, filters, kernel size, dropout, learning rate, and batch size change the result.

Grid and random search become costly; black-box optimization may hide which parameters matter.

The goal is not a heavier network; the goal is a better configuration under a constrained budget.

U-Net

+

Hyperparameters

Lightweight search

ICCS 2026 | DESY, Hamburg | 29 June - 1 July 2026

02

3 of 10

Core idea: GWO tunes a parameterized U-Net

Each wolf represents one candidate hyperparameter vector.

Candidate vector

x = [η, p, F, K, B, D]

η: learning rate

p: dropout rate

F, K, B, D: base filters, kernel, batch, depth

Fitness minimizes segmentation error:

f(x) = 1 - DSC

ICCS 2026 | DESY, Hamburg | 29 June - 1 July 2026

03

4 of 10

Lightweight optimization workflow

Short, low-resolution evaluations first; full-resolution retraining only at the end.

1

Initialize wolves

sample search space

2

Evaluate quickly

downscaled inputs

3

Update leaders

α, β, δ guide search

4

Select best

stable configuration

5

Retrain

full-resolution model

Search range

η: 1e-5 to 1e-2

p: 0.0 to 0.5

F: 16/32/48/64

K: 3 or 5

B: 4/8/16

D: 2/3/4

The search is intentionally compact: optimize only the parameters that change capacity, receptive field, regularization, and training dynamics.

ICCS 2026 | DESY, Hamburg | 29 June - 1 July 2026

04

5 of 10

Experimental setup

Two datasets, compact preprocessing, and overlap/boundary evaluation.

FBTS

T1-CE MRI slices

Meningioma • Glioma • Pituitary

Binary masks

BraTS 2021

FLAIR • T1 • T1CE • T2

WT / TC / ET annotations

Generalization test

Preprocessing

grayscale conversion

resize and min-max normalization

short optimization training, then final retraining

Metrics

DSC

JI

HD

ASSD

Training: Adam optimizer with binary cross-entropy; random seeds fixed.

ICCS 2026 | DESY, Hamburg | 29 June - 1 July 2026

05

6 of 10

Optimization behavior: fast convergence, interpretable sensitivity

The search finds a stable region rather than a single isolated optimum.

~2

iterations to jump from ≈0.26 to >0.82 Dice

≈10

evaluations to exceed 0.80 best-so-far Dice

10⁻³

learning-rate region with strongest performance

D=3

moderate depth provides stable performance

ICCS 2026 | DESY, Hamburg | 29 June - 1 July 2026

06

7 of 10

Segmentation performance

High FBTS cross-validation scores and strong BraTS 2021 whole-tumor generalization.

Best reported values

FBTS DSC 0.9820

FBTS JI 0.9652

BraTS WT DSC 0.9778

BraTS WT JI 0.9566

ICCS 2026 | DESY, Hamburg | 29 June - 1 July 2026

07

8 of 10

Comparison and qualitative evidence

The optimized model is competitive while remaining lightweight.

Visual masks show strong overlap with ground truth; boundary errors are mostly localized around ambiguous tumor edges.

FBTS DSC improves over representative U-Net-based methods.

The method tunes the existing U-Net rather than adding a large architecture.

ICCS 2026 | DESY, Hamburg | 29 June - 1 July 2026

08

9 of 10

What the paper contributes

A compact optimization study, not only a final score.

1

Lightweight GWO tuning

small population and short evaluations

2

Parameterized U-Net

architecture + training variables

3

Behavior analysis

convergence, sensitivity, Top-K region

4

Generalization check

FBTS and BraTS 2021 evaluation

Key message

Metaheuristic tuning can improve segmentation quality while keeping the final U-Net compact and interpretable.

ICCS 2026 | DESY, Hamburg | 29 June - 1 July 2026

09

10 of 10

Take-home message

A lightweight GWO search can identify high-performing U-Net settings under limited computational budget.

Best FBTS score: DSC 0.9820.

Best BraTS 2021 whole-tumor score: DSC 0.9778.

The most informative parameter was the learning rate; moderate depth was most stable.

Acknowledgement

ACK Cyfronet AGH

Grant no. PLG/2025/018784

Polish Ministry of Science and Higher Education funds assigned to AGH University of Krakow

Thank you

Questions?