An enhanced U-Net framework with deep convolutional block optimization for accurate pain state detection in stiff knee joints

After total knee arthroplasty (TKA), knee stiffness is a very uncommon but extremely complicated condition that poses serious problems for both pathophysiology and therapeutic management. In order to break intra-articular adhesions, the knee joint is moved toward the maximum attainable flexion angle (maxAP) during rehabilitation; nevertheless, this procedure frequently causes significant discomfort and, if too much power is used, may result in subsequent injury. This study uses discriminative characteristics extracted from electromyography (EMG) data to estimate postoperative maxAP in order to provide objective and safe measurement of maxAP. As a multivariate pain-state classification problem, the maxAP detection job is designed to meet clinical standards. To address this classification issue, we suggest the Deep Convolution Block Optimized U-Net (DCBO-UN) model. The suggested model, in contrast to the conventional U-Net, uses deeper convolutional blocks to improve representation of minor EMG fluctuations linked to pain states and to improve hierarchical feature extraction. In order to adaptively recalibrate channel-wise and spatial feature responses, block attention techniques are also incorporated into both the downsampling and upsampling paths. This allows the network to suppress redundant activations and noise while concentrating on therapeutically relevant signal components.When compared to the traditional U-Net framework, DCBO-UN improves feature discrimination and contextual awareness through these architectural improvements. The suggested model’s improved classification performance in maxAP estimate is demonstrated by experimental results, underscoring its potential to support robotic-assisted rehabilitation programs and enhance the accuracy and safety of postoperative knee stiffness management.