Optimizing Neural Network Architecture for Medical Image Segmentation Using Monte Carlo Tree Search

Liping Meng; Fan Nie; Yunyun Zhang; Chao Han

cs.CVFeb 25, 2026

Optimizing Neural Network Architecture for Medical Image Segmentation Using Monte Carlo Tree Search

Authors: Liping Meng, Fan Nie, Yunyun Zhang, Chao Han

Abstract

This paper proposes a novel medical image segmentation framework, MNAS-Unet, which combines Monte Carlo Tree Search (MCTS) and Neural Architecture Search (NAS). MNAS-Unet dynamically explores promising network architectures through MCTS, significantly enhancing the efficiency and accuracy of architecture search. It also optimizes the DownSC and UpSC unit structures, enabling fast and precise model adjustments. Experimental results demonstrate that MNAS-Unet outperforms NAS-Unet and other state-of-the-art models in segmentation accuracy on several medical image datasets, including PROMISE12, Ultrasound Nerve, and CHAOS. Furthermore, compared with NAS-Unet, MNAS-Unet reduces the architecture search budget by 54% (early stopping at 139 epochs versus 300 epochs under the same search setting), while achieving a lightweight model with only 0.6M parameters and lower GPU memory consumption, which further improves its practical applicability. These results suggest that MNAS-Unet can improve search efficiency while maintaining competitive segmentation accuracy under practical resource constraints.

Figures & tables

Figure 3: (a) Architecture of U-Net. (b) The U-shaped structure of the MNAS-Unet is depicted with rectangles, each representing a cell architecture subject to optimization. The specific operations of DownSC and UpSC are integrated into MNAS-Unet, corresponding to downward and upward movements, respectively. A blue arrow indicates the progression of the feature map (input image). Additionally, the light purple arrow indicates a transformation operation within UpSC that is also subject to automatic search. Adapted from [ 10 ] .

POs type	Down POs	Up POs	Normal POs
1	avg pool	up cweight	identity
2	max pool	up dep conv	dep conv
3	down conv	up conv	dil conv
4	down dep conv	up dil conv	cweight
5	down dil conv	–	conv
6	down cweight	–	shuffle conv

Table 1: Search space for MNAS-Unet architecture search. Operations are selected based on genotype analysis from experimental results. The search space includes 6 Down POs, 4 Up POs, and 6 Normal POs.

Model

PROMISE12

Ultrasound

CHAOS

mIoU

DSC

Time

mIoU

DSC

Time

mIoU

DSC

Time

U-Net

0.515

0.680

46min

6.5GB

0.662

0.797

6h05m

7.0GB

0.886

0.939

13h08m

8.0GB

FC-DenseNet

0.528

0.691

53min

7.2GB

0.673

0.805

6h28m

7.4GB

0.902

0.948

13h22m

8.0GB

DC-UNet

0.533

0.695

49min

6.5GB

0.676

0.808

6h02m

6.9GB

0.898

0.946

13h05m

7.5GB

DCSAU-Net

0.539

0.701

51min

6.8GB

0.681

0.811

6h14m

7.0GB

0.905

0.952

13h12m

7.7GB

MNet

0.534

0.696

48min

6.2GB

0.684

0.813

5h47m

6.6GB

0.907

0.953

12h34m

7.3GB

Table 2: Performance comparison across datasets. MNAS-Unet achieves the best overall segmentation performance while maintaining lower memory consumption. Training time refers to single-run wall-clock time measured under identical batch size and input resolution.

Explore similar work

Active-DiNTS: Active Differentiable Network Topology Search
Oct 3, 2026Gean Trindade Pereira, Thierry Urruty, Muriel Visani +1Neural Architecture Search 3D Medical Image Segmentation
Institute of Mathematics and Computer Sciences (ICMC) University of São Paulo São Carlos, 13560-970, São Paulo, Brazil · XLIM, Université de Poitiers Université de Limoges, CNRS Poitiers, 86000, Vienne, France · Laboratoire Informatique, Image et Interaction (L3i) La Rochelle Université La Rochelle, 17042, Charente-Maritime, France
Resource-Aware Evolutionary Neural Architecture Search for Cardiac MRI Segmentation
May 7, 2026Farhana Yasmin, Mahade Hasan, Haipeng Liu +3Medical Image Analysis Medical Image Segmentation
School of Computer Science, Nanjing University of Information Science and Technology, Nanjing, China. · 3Eastern University, Ashulia Model Town, Dhaka, 1345, Bangladesh. · School of Software, Nanjing University of Information Science and Technology, Nanjing, China.
XTinyU-Net: Training-Free U-Net Scaling via Initialization-Time Sensitivity
May 10, 2026Alvin Kimbowa, Moein Heidari, David Liu +1Model Compression U-Net
School of Biomedical Engineering, The University of British Columbia, Vancouver, Canada · Department of Radiology, The University of British Columbia, Vancouver, Canada · Department of Medicine, The University of British Columbia, Vancouver, Canada

Oct 3, 2026J/K move · Enter open · S save

Active-DiNTS: Active Differentiable Network Topology Search

Gean Trindade Pereira, Thierry Urruty, Muriel Visani, André C. P. L. F. de Carvalho

Institute of Mathematics and Computer Sciences (ICMC) University of São Paulo São Carlos, 13560-970, São Paulo, Brazil · XLIM, Université de Poitiers Université de Limoges, CNRS Poitiers, 86000, Vienne, France · Laboratoire Informatique, Image et Interaction (L3i) La Rochelle Université La Rochelle, 17042, Charente-Maritime, France

Neural Architecture Search 3D Medical Image Segmentation Active Learning Brain Tumor Segmentation

Open paper AskPDF arXiv

Neural Architecture Search (NAS) has proved to be a strong alternative to manual network design, but applying it to 3D medical image segmentation is limited by two well-known costs, large annotation budgets and multi-GPU clusters. Thus, this paper introduces Active-DiNTS (Active Differentiable Network Topology Search), an approach that embeds pool-based Active Learning (AL) into a bi-level differentiable topology search to perform architecture discovery and label curation jointly. At each query round, unlabeled MRI volumes are ranked by one of three uncertainty signals (Entropy, Variance, or Standard Deviation), and only the top-ranked volumes are sent to an oracle for annotation. The new labels feed two interlocked stages. Network weights are updated in an outer loop, while the macro/micro topology of a U-Net-style backbone is refined in an inner loop. Three AL regimes (weights-only, topology-only, joint) expose the speed-accuracy trade-off. Evaluations on the Medical Segmentation Decathlon (MSD) Task01 BrainTumour benchmark showed that Active-DiNTS surpasses DiNTS, C2FNAS, and nnU-Net in Dice, with gains of about 10 percentage points on Edema and 5 points on Non-Enhancing core, on a single GPU and using a fraction of the labeled volumes. The discovered architectures are denser and more FLOP-heavy than prior baselines, but remain competitive in trainable parameters and peak memory; the fastest search variant finishes in under 0.25 GPU-days, over 27x faster than the eight-GPU DiNTS search. Together, these results indicate that pairing differentiable NAS with active data acquisition is a practical recipe for accurate 3D segmentation under realistic constraints.

Abstract

Explore similar work

Active-DiNTS: Active Differentiable Network Topology Search

Resource-Aware Evolutionary Neural Architecture Search for Cardiac MRI Segmentation

XTinyU-Net: Training-Free U-Net Scaling via Initialization-Time Sensitivity