Exploiting Architectural and Loss Complementarity: A Diversity-Driven Ensemble for Brain Tumor Segmentation on BraTS 2020

Author(s)	Mr. S. U. Ravi Kumar Chavali, Dr. V. Kumar P
Country	India
Abstract	Deep-learning brain-tumor segmentation on BraTS is dominated by U-Net-family models, yet a prior matched-condition study established that no single architecture or loss dominates and that different configurations excel on different tumor sub-regions. It is asked whether this complementarity is an exploitable resource. Using sixteen previously trained U-Net-family configurations (four decoders × four losses, sharing a ResNet34 encoder), softmax-averaging ensembles are formed and evaluated on 2,242 BraTS 2020 test slices, with test-time augmentation applied uniformly so that gains reflect ensembling alone. A sixteen-model ensemble raises mean Dice from 0.831 to 0.842 (Δ = +0.011, 95% CI [+0.007, +0.015], p < 0.001), with the largest gain in the tumor-core region. A three-model complementary subset recovers most of the benefit, and both architectural and loss diversity contribute significantly. Ensembling converts matched-comparison complementarity into a small but statistically robust accuracy gain at modest additional cost.
Keywords	Brain Tumor Segmentation, BraTS 2020, Deep Ensemble, Model Complementarity, Test-Time Augmentation
Field	Computer > Artificial Intelligence / Simulation / Virtual Reality
Published In	Volume 8, Issue 4, July-August 2026
Published On	2026-08-10
DOI	https://doi.org/10.36948/ijfmr.2026.v08i04.85534

About IJFMR Fees & Payment Current Issue Publication Archive	Submit Research Paper Track Submission Status Publication Guidelines Publication Ethics Peer Review & Plagiarism	Join as a Reviewer Editors & Reviewers Reviewer Referral Program Get Reviewer Membership Certi.	Website/Journal Policies Usage Policy Content Policies Privacy Policy

Contact Us		+91-9687-828-838	editor@ijfmr.com

International Journal For Multidisciplinary Research