Body Composition Analysis
Introduction
Advances in deep learning-based computer vision technologies continue to evolve, from CNNs to Vision Transformers. This study focuses on the development of a CNN-Transformer model for body composition analysis using CT image segmentation and explores semi-supervised learning techniques to address the shortage of labeled medical images. By extracting health-related information and facilitating medical decision-making, this research aims to utilize these insights for medical applications.
In addition to the prior studies described below, we are conducting various deep learning-based medical studies, including CT-based 3D body composition analysis, multi-modal fusion techniques combining EMR(text) and CT(images), and early cancer detection models.
Abdominal CT Segmentation for Body Composition Assessment Using Network Consistency Learning
Estimating skeletal muscle (SM) and adipose tissues is an invaluable prognostic indicator in cancer treatment, major surgeries, and general health screening. Body composition is usually measured with abdominal computed tomography (CT) scans acquired in clinical settings. The whole-body SM volume is correlated with the estimated SM based on the measurement of a single two-dimensional vertebral slice. It is necessary to label a CT image at the pixel level to estimate SM, known as semantic segmentation. In this work, we trained a segmentation model using the labeled abdominal CT slices and the additional unlabeled slices. In particular, we trained two identical segmentation networks with differently initialized weights. Network Consistency Learning (NCL) allowed learning from unlabeled images by forcing the predictions from both networks to be the same. We segmented abdominal CT images from a newly created in-house dataset. The proposed approach gained 10% better performance in terms of Dice similarity score (DSC) than that obtained by a standard supervised network demonstrating the effectiveness of NCL in exploiting unlabeled images.
Fuzzy boundary based Downsampling and Class-aware Downsampling
Convolutional operations in convolutional neural networks (CNNs) require significant memory usage. Training a segmentation model in semi-supervised settings to leverage unlabeled CT images can take days to weeks. One potential approach to expedite this process is to utilize lower-resolution CT images to decrease spatial dimensions. However, downsampling inevitably leads to information loss – a drawback that must be addressed.
We aim to utilize PixelShuffle and PixelUnShuffle to decrease the spatial size while increasing the channel dimensions to address information loss. Performing convolutional operations on smaller images reduces computational requirements, resulting in quicker training times and reduced memory usage. Moreover, training with multiple channels can offer the model additional information.
In medical images, where each pixel's contribution is significant, our research aims to develop a class-aware probabilistic method for encoding class uncertainty during the downsampling of segmentation masks. This approach transform segmentation masks with hard labels (discrete one-hot encoding) into soft labels (continuous values). Additionally, we plan to thoroughly investigate potential performance improvements and develop similar downsampling methods for CT images to retain as much information as possible.
L3 Slice Detector in CT Scans Multi-Channel Inputs
A comprehensive analysis of the L3 lumbar spine region is crucial for evaluating body composition and discerning the distribution of muscle and fat tissue. Muscle mass or fat distribution changes 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. The primary objective of this work is to develop an L3 slice detector to classify slices in a given CT scan, either as L3 or non-L3. The proposed L3 slice detector has minimal parameters and uses the multi-channel approach for the dataset.
Papers
Iftikhar Ahmad, Shahzad Ali, Yu Rim Lee, Soo Young Park, Won Young Tak and Soon Ki Jung, L3 Slice Detection in CT Scans via Pixel-Unshuffled Multi-Channel Inputs, The International Workshop on Frontiers of Computer Vision (IW-FCV), (2024.02.19 ~ 2024.02.21)
Shahzad Ali, Yu Rim Lee, Soo Young Park, Won Young Tak, and Soon Ki Jung, Volumetric Body Composition through Cross-Domain Consistency Training for Unsupervised Domain Adaptation, The 18th International Symposium on Visual Computing (ISVC) (2023.10.16 ~ 2023.10.18)
Shahzad Ali, Yu Rim Lee, Soo Young Park, Won Young Tak, Soon Ki Jung, Abdominal CT Segmentation for Body Composition Assessment using Network Consistency Learning, The 45th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC 2023) (2023.07.24 ~ 2023.07.27)
Muhammad Salman Khan, Shahzad Ali (equally contributed), Yu Rim Lee, Soo Young Park, Won Young Tak, Soon Ki Jung, Cell Nuclei Segmentation With Dynamic Token-Based Attention Network, the 45th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC2023) (2023.07.24 ~ 2023.07.27)
Muhammad Salman Khan, Shahzad Ali, Yu Rim Lee, Min Kyu Kang, Soo Young Park, Won Young Tak, and Soon Ki Jung, TransUNet-Lite: A Robust Approach to Cell Nuclei Segmentation, The 7th International Conference on Medical and Health Informatics (ICMHI 2023) (2023.05.12 ~ 2023.05.14).
Shahzad Ali, Arif Mahmood, and Soon Ki Jung, Lightweight Encoder-Decoder Architecture for Foot Ulcer Segmentation, The 28th International Workshop on Frontiers of Computer Vision(IW-FCV 2022), 2022 (2022.02.21 ~ 2022.02.22)