Liver Fat Quantification Network with Body Shape

TL;DR

The paper addresses the problem of non-invasively quantifying liver fat percentage from body shape, highlighting the inadequacy of grade-based assessment for continuous monitoring. It introduces a two-stream regression network with a lightweight multi-channel attention module that maps frontal and lateral 2D body shape maps to liver fat percentage, leveraging joint regression and classification objectives. On a CT-derived dataset with ground-truth liver fat from HU averages, the method achieves a regression RMSE of $5.26\%$ and $R^2$ of $0.815$, outperforming linear models and baseline deep-learning architectures. The approach offers a cheaper, more accessible alternative for monitoring hepatic steatosis, with interpretable region-based activations via Grad-CAM.

Abstract

It is critically important to detect the content of liver fat as it is related to cardiac complications and cardiovascular disease mortality. However, existing methods are either associated with high cost and/or medical complications (e.g., liver biopsy, imaging technology) or only roughly estimate the grades of steatosis. In this paper, we propose a deep neural network to estimate the percentage of liver fat using only body shapes. The proposed is composed of a flexible baseline network and a lightweight Attention module. The attention module is trained to generate discriminative and diverse features which significant improve the performance. In order to validate the method, we perform extensive tests on the public medical dataset. The results verify that our proposed method yields state-of-the-art performance with Root mean squared error (RMSE) of 5.26% and R-Squared value over 0.8. It offers an accurate and more accessible assessment of hepatic steatosis.

Figures (7)

• Figure 1: The extraction flowchart of 2D body shape maps and liver fat. The original CTs are calibrated and aligned to the same standard with the pipeline. The frontal and lateral shape maps are extracted from each slice are then combined together. The percentage of liver fat and hepatic steatosis degrees are used as the ground truth is estimated by averaging HU values of different liver regions.
• Figure 2: The distribution of steatosis grades. The number of each steatosis grades (grade 0, 1, 2, 3) are 122, 107, 42, 44.
• Figure 3: The architecture of the proposed network. The network includes a baseline regression network and a multi-channel attention module. The input of the network is two body shape maps frontal and lateral body shape maps, the outputs of network are liver fat percentage and steatosis grades.
• Figure 4: The detail structure of attention module. The attention module owns four channel whihc is consistent with the steatosis grades. It consists of two component: classification and regression component.
• Figure 5: The predicted and truth liver fat percentages results of different methods. From (a) - (f) are results of Linear Regresssion, SVM, MLP, ResNet, Baseline and proposed method.