NeuroMemory-DiffNet: a reference-guided memory framework for brain tumor segmentation and classification

The diagnosis of brain tumors is still difficult due to the diversity of tumor morphology, inter-individual anatomical differences and the limited interpretability of traditional deep learning models, despite the progress of magnetic resonance imaging (MRI) technology. In order to overcome these challenges, this paper introduces a novel explainable framework named NeuroMemory-DiffNet (NMD-Net), which combines a Neuro-Anatomical Prior Memory (NAPM), Reference-Guided Difference Learning (RDL) and Healthy-Reference Reconstruction (HRR) to support anatomically guided brain tumor diagnosis. NMD-Net accesses a reference neuro-anatomical representation (NAR) derived from training, and uses RDL to estimate deviations between the observed MRI representation and the retrieved representation, which are related to the disease. The HRR branch is used to obtain a reference reconstruction and residual representation, highlighting the structural deviation toward the disease and minimising the impact of normal anatomical variation. To ensure structural coherence and facilitate strong feature learning and model level interpretability, an anatomical consistency constraint is added. The framework was tested on BraTS 2023 in the task of glioma segmentation and on the Figshare-based brain tumor dataset in the task of multi-class classification. NeuroMemory-DiffNet obtained an overall accuracy of 98.70%, a macro-averaged precision of 98.71%, a macro-averaged sensitivity of 98.70%, a macro-averaged specificity of 99.57% and a macro-averaged F1-score of 98.70% on the subject-independent test dataset for multi-class brain tumor classification. For tumor segmentation, the framework achieved a Dice Similarity Coefficient (DSC) of 96.31%, Intersection over Union (IoU) of 93.62%, and Hausdorff Distance (HD95) of 2.29 mm and a false-positive rate (FPR) of 0.94%. Competitive performance to the benchmark level models evaluated in this work was achieved when compared with the evaluated models. The results show the benchmark level performance of NMD-Net for brain tumor classification and segmentation which also serve as model-level reference-guided explanations; however, they require further external and expert validation before making them clinically available.

Authors

Institutions

Publication Details

Journal
Scientific Reports
Published
2026-11-12
DOI
https://doi.org/10.1038/s41598-026-79184-9
Primary Topic
Brain Tumor Detection and Classification
Type
article
Field-Weighted Citation Impact
0.00

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

NeuroMemory-DiffNet: a reference-guided memory framework for brain tumor segmentation and classification

Parameswari Alagarsamy, Ayodele Lasisi, Temesgen Engida Yimer, Rashmi Hegde et al.
Scientific Reports
Brain Tumor Detection and Classification
article

NeuroMemory-DiffNet: a reference-guided memory framework for brain tumor segmentation and classification

Parameswari Alagarsamy, Ayodele Lasisi, Temesgen Engida Yimer, Rashmi Hegde, B. R. Chandana, Saiful Islam, Quadri Noorulhasan Naveed
article en

Abstract

The diagnosis of brain tumors is still difficult due to the diversity of tumor morphology, inter-individual anatomical differences and the limited interpretability of traditional deep learning models, despite the progress of magnetic resonance imaging (MRI) technology. In order to overcome these challenges, this paper introduces a novel explainable framework named NeuroMemory-DiffNet (NMD-Net), which combines a Neuro-Anatomical Prior Memory (NAPM), Reference-Guided Difference Learning (RDL) and Healthy-Reference Reconstruction (HRR) to support anatomically guided brain tumor diagnosis. NMD-Net accesses a reference neuro-anatomical representation (NAR) derived from training, and uses RDL to estimate deviations between the observed MRI representation and the retrieved representation, which are related to the disease. The HRR branch is used to obtain a reference reconstruction and residual representation, highlighting the structural deviation toward the disease and minimising the impact of normal anatomical variation. To ensure structural coherence and facilitate strong feature learning and model level interpretability, an anatomical consistency constraint is added. The framework was tested on BraTS 2023 in the task of glioma segmentation and on the Figshare-based brain tumor dataset in the task of multi-class classification. NeuroMemory-DiffNet obtained an overall accuracy of 98.70%, a macro-averaged precision of 98.71%, a macro-averaged sensitivity of 98.70%, a macro-averaged specificity of 99.57% and a macro-averaged F1-score of 98.70% on the subject-independent test dataset for multi-class brain tumor classification. For tumor segmentation, the framework achieved a Dice Similarity Coefficient (DSC) of 96.31%, Intersection over Union (IoU) of 93.62%, and Hausdorff Distance (HD95) of 2.29 mm and a false-positive rate (FPR) of 0.94%. Competitive performance to the benchmark level models evaluated in this work was achieved when compared with the evaluated models. The results show the benchmark level performance of NMD-Net for brain tumor classification and segmentation which also serve as model-level reference-guided explanations; however, they require further external and expert validation before making them clinically available.

Scientific Reports
Nitte University (IN), Dilla University (ET), Hindustan Institute of Technology and Science (IN), Sahyadri College of Engineering & Management (IN), King Khalid University (SA)
King Khalid University
Good health and well-being
Brain Tumor Detection and Classification
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon