Pixel-Unshuffled Multi-Channel Approach for Efficient L3 Slice Detection in CT Scans

초록

A comprehensive analysis of the L3 lumbar spine region is crucial for evaluating body composition and understanding the distribution of muscle and fat tissue. Changes in muscle mass or fat distribution are linked to several disorders that significantly impact quality of life. Therefore, there is a need for a fully automated system to extract an L3 slice, enabling efficient and precise analysis. This study aims to develop an L3 slice detector that classifies slices in a given CT scan as either L3 or non-L3. The proposed L3 slice detector is designed with minimal parameters and utilizes a multi-channel approach for the dataset. The multi-channel inputs comprise various slices across several channels, including both L3 and non-L3 slices. A significant challenge in identifying the L3 slice is the limited knowledge of its exact location. For each patient, a single slice is selected as L3 from several candidates, resulting in a relatively low number of L3 slices available for training the network. This scarcity creates an imbalance in the dataset, reducing classification performance when trained on single-slice input. A multi-channel input approach addresses this issue and enhances the detector's performance. The slice detection network is based on a convolutional neural network (CNN), which classifies an input as L3 or non-L3. We used an in-house dataset of CT (Computed Tomography) volumes with marked L3 locations for experiments. We conducted two experiments, one with single-channel inputs and another with multi-channel inputs. The multi-channel approach yielded better results than the single-channel input, achieving an F1 score of 86.49. © 2024 Copyright held by the owner/author(s).

키워드