A novel negative-based CNN and deep feature engineering framework for ACL rupture detection from knee MRI
Abstract
Deep learning is widely used in biomedical image classification, but many studies rely on established architectures. This study presents NegativeKANNeXt, a compact convolutional architecture for anterior cruciate ligament (ACL) rupture detection from knee magnetic resonance imaging (MRI), together with a deep feature engineering (DFE) pipeline. NegativeKANNeXt uses subtraction-based and additive feature fusion and applies the Gaussian Error Linear Unit (GELU) and Swish functions to shared pre-activations. The dual-activation design is inspired by the Kolmogorov-Arnold Network (KAN) formulation, but it is not equivalent to KAN. The network contains Negative-Stem, NegativeKAN, Negative-Down, and a conventional output phase. In the DFE pipeline, global average pooling features from the trained network are reduced by CWINCA and classified by shallow classifiers. The cohort contains 681 patients, with one representative sagittal slice per patient. Patient-level partitioning was completed before augmentation, and augmentation was confined to the training partition. Evaluation included a patient-level hold-out test, repeated patient-level cross-validation, and frozen-model external validation. In repeated cross-validation, NegativeKANNeXt achieved 93.54 ± 2.32% accuracy, whereas DFE + SVM achieved 95.75 ± 2.32%. On the held-out, non-augmented test partition, the corresponding accuracies were 93.38% and 96.32%, with four selected features for DFE + SVM. Frozen-model external validation yielded 83.33% accuracy under domain shift. These results support subtraction-based fusion as a viable compact CNN design. The DFE result applies to the NegativeKANNeXt-based pipeline; equivalent DFE pipelines were not evaluated for the baseline backbones.
Related
- Cross-corpus consensus feature selection for language-invariant Parkinson’s disease detection: a dual-stream acoustic and spectrogram-attention framework
- Development of a parallel cascaded adaptive ensemble framework with feature engineering process for the prediction of neurodegenerative disease using temporal clinical data
- Estimating Heart Failure Health Status From Multimodal Wearables: Implementation of a Feature-Engineering Framework for KCCQ-12 Prediction.
- Multimodality imaging for early detection of subacute myocardial rupture
- Optimized Feature Engineering and Machine Learning Approaches for Early Prediction of Renal Cell Carcinoma
BibTeX
@article{demir2026novel,
title = {{A novel negative-based CNN and deep feature engineering framework for ACL rupture detection from knee MRI}},
author = {Sukru Demir and Bugra Can and Omer Faruk Goktas and Mehmet Baygin and Sengul Dogan and Turker Tuncer},
journal = {Scientific Reports},
year = {2026},
month = oct,
doi = {10.1038/s41598-026-74244-6},
url = {https://www.nature.com/articles/s41598-026-74244-6},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - A novel negative-based CNN and deep feature engineering framework for ACL rupture detection from knee MRI
AU - Sukru Demir
AU - Bugra Can
AU - Omer Faruk Goktas
AU - Mehmet Baygin
AU - Sengul Dogan
AU - Turker Tuncer
PY - 2026
DA - 2026/10/07
JO - Scientific Reports
DO - 10.1038/s41598-026-74244-6
UR - https://www.nature.com/articles/s41598-026-74244-6
AB - Deep learning is widely used in biomedical image classification, but many studies rely on established architectures. This study presents NegativeKANNeXt, a compact convolutional architecture for anterior cruciate ligament (ACL) rupture detection from knee magnetic resonance imaging (MRI), together with a deep feature engineering (DFE) pipeline. NegativeKANNeXt uses subtraction-based and additive feature fusion and applies the Gaussian Error Linear Unit (GELU) and Swish functions to shared pre-activations. The dual-activation design is inspired by the Kolmogorov-Arnold Network (KAN) formulation, but it is not equivalent to KAN. The network contains Negative-Stem, NegativeKAN, Negative-Down, and a conventional output phase. In the DFE pipeline, global average pooling features from the trained network are reduced by CWINCA and classified by shallow classifiers. The cohort contains 681 patients, with one representative sagittal slice per patient. Patient-level partitioning was completed before augmentation, and augmentation was confined to the training partition. Evaluation included a patient-level hold-out test, repeated patient-level cross-validation, and frozen-model external validation. In repeated cross-validation, NegativeKANNeXt achieved 93.54 ± 2.32% accuracy, whereas DFE + SVM achieved 95.75 ± 2.32%. On the held-out, non-augmented test partition, the corresponding accuracies were 93.38% and 96.32%, with four selected features for DFE + SVM. Frozen-model external validation yielded 83.33% accuracy under domain shift. These results support subtraction-based fusion as a viable compact CNN design. The DFE result applies to the NegativeKANNeXt-based pipeline; equivalent DFE pipelines were not evaluated for the baseline backbones.
SN - 2045-2322
ER -