Multi-scale convolution with efficient channel attention improves benign–malignant classification of spinal tumors on T2-weighted fat-suppressed MRI
Abstract
BackgroundAccurate differentiation between benign and malignant spinal tumors is critical for treatment planning, yet manual interpretation of T2-weighted fat-suppressed MRI remains subjective, with substantial inter-observer variability. This study aimed to develop a deep learning model, the multi-scale deep residual network (MSDR-Net), for automated benign–malignant classification.MethodsA total of 500 T2-weighted fat-suppressed MR images of pathologically confirmed spinal tumors (389 benign, 111 malignant) were retrospectively collected from two institutions. Data were partitioned into training (70%), validation (20%), and independent test (10%) sets, and five-fold cross-validation was performed. Leave-one-center-out external validation was also performed (development: Institution A, n = 300; evaluation: Institution B, n = 200). The proposed MSDR-Net incorporates multi-branch residual blocks that combine parallel multi-scale convolutional paths with efficient channel attention modules and was benchmarked against six established architectures.ResultsOn the independent test set (N = 50; 12 malignant cases), MSDR-Net achieved 92.0% accuracy, 91.7% sensitivity (95% CI 61.5%–99.8%), 92.1% specificity, and an AUC of 0.921, the highest point estimates among evaluated architectures. The model was well calibrated (Brier score 0.072), with net benefit over treat-all strategies across thresholds of 0.10–0.60. Derived from a small test cohort, these estimates carry substantial uncertainty and await confirmation in larger, prospective cohorts. External-cohort accuracy was 86.5% (AUC 0.875). Ablation experiments showed that removing either component degraded performance, with a modest super-additive interaction; given their substantially smaller capacity, the single-path variants’ deficit cannot be attributed solely to removing multi-scale processing. The model has 55.4M parameters and 12.3-ms inference per image.ConclusionMSDR-Net effectively captures multi-scale discriminative features of spinal tumors on T2-weighted fat-suppressed MRI, achieving high classification accuracy while maintaining computational efficiency. The model shows promise for clinical deployment as an auxiliary diagnostic tool for orthopedic surgeons and radiologists.
Related
- Feature-aware attention convolutional neural network for efficient pulmonary fibrosis classification from chest CT images with FPGA deployment
- Multi-view attention-based deep learning for benign–malignant classification of pulmonary nodules
- TiLCA-Net: nuclei segmentation and classification in pathological images via task-specialized attention and efficient decoding
- HayabusaNet: Hybrid attention-based multiscale fusion CNN for accurate and efficient brain tumor classification in MRI scans
- Asymmetric encoder-decoder U-net with hybrid attention and residual dilated convolution for brain tumor segmentation
BibTeX
@article{ma2026multiscale,
title = {{Multi-scale convolution with efficient channel attention improves benign–malignant classification of spinal tumors on T2-weighted fat-suppressed MRI}},
author = {Zhiwei Ma and Xingyu Duan and Xiaorong Ma and Cuiping Yan and Xuewei Wang and Ningkui Niu},
journal = {Frontiers in Oncology},
year = {2026},
month = oct,
doi = {10.3389/fonc.2026.1951483},
url = {https://www.frontiersin.org/articles/10.3389/fonc.2026.1951483},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Multi-scale convolution with efficient channel attention improves benign–malignant classification of spinal tumors on T2-weighted fat-suppressed MRI
AU - Zhiwei Ma
AU - Xingyu Duan
AU - Xiaorong Ma
AU - Cuiping Yan
AU - Xuewei Wang
AU - Ningkui Niu
PY - 2026
DA - 2026/10/30
JO - Frontiers in Oncology
DO - 10.3389/fonc.2026.1951483
UR - https://www.frontiersin.org/articles/10.3389/fonc.2026.1951483
AB - BackgroundAccurate differentiation between benign and malignant spinal tumors is critical for treatment planning, yet manual interpretation of T2-weighted fat-suppressed MRI remains subjective, with substantial inter-observer variability. This study aimed to develop a deep learning model, the multi-scale deep residual network (MSDR-Net), for automated benign–malignant classification.MethodsA total of 500 T2-weighted fat-suppressed MR images of pathologically confirmed spinal tumors (389 benign, 111 malignant) were retrospectively collected from two institutions. Data were partitioned into training (70%), validation (20%), and independent test (10%) sets, and five-fold cross-validation was performed. Leave-one-center-out external validation was also performed (development: Institution A, n = 300; evaluation: Institution B, n = 200). The proposed MSDR-Net incorporates multi-branch residual blocks that combine parallel multi-scale convolutional paths with efficient channel attention modules and was benchmarked against six established architectures.ResultsOn the independent test set (N = 50; 12 malignant cases), MSDR-Net achieved 92.0% accuracy, 91.7% sensitivity (95% CI 61.5%–99.8%), 92.1% specificity, and an AUC of 0.921, the highest point estimates among evaluated architectures. The model was well calibrated (Brier score 0.072), with net benefit over treat-all strategies across thresholds of 0.10–0.60. Derived from a small test cohort, these estimates carry substantial uncertainty and await confirmation in larger, prospective cohorts. External-cohort accuracy was 86.5% (AUC 0.875). Ablation experiments showed that removing either component degraded performance, with a modest super-additive interaction; given their substantially smaller capacity, the single-path variants’ deficit cannot be attributed solely to removing multi-scale processing. The model has 55.4M parameters and 12.3-ms inference per image.ConclusionMSDR-Net effectively captures multi-scale discriminative features of spinal tumors on T2-weighted fat-suppressed MRI, achieving high classification accuracy while maintaining computational efficiency. The model shows promise for clinical deployment as an auxiliary diagnostic tool for orthopedic surgeons and radiologists.
SN - 2234-943X
ER -