MEFA-Unet: Multi-scale feature extraction and fusion attentional unet for segmenting short process of incus in...

The application of convolutional neural networks (CNNs) in microsurgery has been limited. Current CNNs struggle to capture diverse semantic details across various scales and receptive fields, and they also face challenges in establishing meaningful connections between features detected by different receptive fields. Thus, they fail to deal with small objects and maintain accurate boundaries, especially in some complex microsurgical scenarios.

We propose a multi-scale feature extraction and fusion attentional Unet (MEFA-Unet) to address these limitations. This model employs a hybrid attention mechanism integrating Squeeze-and-Excitation (SE) and Coordinate Attention (CA) modules to enhance feature focus. The MEFA-Unet utilizes multi-scale feature fusion at skip connections and across a multi-layer decoding stage, preserving detailed and contextual information for precise image analysis. For training and validation, the dataset comprising 33,420 human-annotated images from 181 patients was constructed based on four microsurgical scenarios. In addition, 1,500 human-annotated images from 30 raw patients and 2 complex video clips were used to test the networks' performance.

Compared to other classical segmentation networks, the MEFA-Unet displays superior segmentation performance in visualization results and evaluation metrics. The proposed method achieved a mean intersection over the unit (mIOU) of 0.8880 with the validation set, a mIOU of 0.7583 with the test set, and a mIOU of 0.7855 with the video. Furthermore, the viability of the trained Unet was confirmed by applying it to real microsurgical procedures, where it exhibited excellent performance even in challenging scenarios characterized by weak texture, low contrast, and varying appearances.

The proposed method exhibits a high level of effectiveness in the automated detection and segmentation of the short process of the incus (SPI) across various microsurgical scenarios.