TurkerNeXtV2: An Innovative CNN Model for Knee Osteoarthritis Pressure Image Classification

Area of Science:

• Computer Vision
• Medical Imaging
• Artificial Intelligence

Background:

• Lightweight Convolutional Neural Networks (CNNs) for medical imaging applications are limited.
• Existing models often struggle to balance effectiveness with computational efficiency.

Purpose of the Study:

• To introduce TurkerNeXtV2, a compact CNN designed for medical imaging.
• To achieve transformer-level effectiveness using CNN simplicity and low computational cost.
• To enhance model stability and efficiency through novel architectural blocks.

Main Methods:

• Developed TurkerNeXtV2 with two new blocks: pooling-based attention with an inverted bottleneck (TNV2) and a hybrid downsampling module.
• Pretrained the model on the Stable ImageNet-1k benchmark.
• Fine-tuned and evaluated on a plantar-pressure osteoarthritis (OA) dataset and a blood-cell image dataset.
• Measured performance using accuracy, precision, recall, F1-score, and inference time (images/second).

Main Results:

• Achieved 87.77% validation accuracy during pretraining on Stable ImageNet-1k.
• Attained 93.40% accuracy on the OA dataset with precision and recall above 90%.
• Reached 98.52% accuracy on the blood-cell dataset.
• Demonstrated an average inference time of 0.0078 seconds per image (≈128.8 images/s), outperforming transformer baselines.

Conclusions:

• TurkerNeXtV2 offers high accuracy and low computational cost for medical imaging tasks.
• The pooling-based attention (TNV2) and hybrid downsampling contribute to a lightweight yet effective design.
• The model is suitable for real-time and clinical deployment.