An Optimized <scp>InceptionNet</scp> Model With Depth‐Wise Separable Convolutions for Cervical Spine Fracture Detection

Abstract

ABSTRACT Cervical spine fractures are serious injuries that can lead to paralysis or death if not detected early. Manual interpretation of medical images, such as CT scans and x‐rays, is time‐consuming and prone to error, especially for subtle fractures. This study proposes a novel deep learning pipeline for automatic cervical spine fracture detection that integrates three key innovations. Initially, input image for pre‐processing using a cross‐guided bilateral filtering (CGBiF) and extended contrast limited adaptive histogram equalization (ECLAHE) to remove unwanted noise and improve local contrast while preserving structure. Then, the bilateral encoder based on U‐Net (BiEU‐Net) is used to perform precise vertebrae segmentation and reduce the computational load. Then, the feature extraction using Convolution Block Encoder Based Autoencoder (CBA‐AE) has both the spatial and channel attention functions that capture the information at multiple levels. Finally, classification using the depth wise separable convolution aided inceptionNet (DS‐INC) has improved the inference speed and high accuracy. The proposed model in public access to the dataset in the RSNA 2022 cervical spine fracture detection, containing 3000 CT studies with interpreted fracture labels. The proposed model achieves 98.37% for the accuracy, 98.35% for the recall, 98.42% for the precision, and 98.38% for the F1‐score for ResNet50, InceptionNet, as well as structures such as Unet++ and swin transformer. This proposed system, due to its high sensitivity, low parameters, and it makes suitable for real‐time clinical application. This work improves the state‐of‐the‐art by combining pre‐processing and classification into a usable system for spinal fracture detection.

Ähnliche Arbeiten