This repository contains the official implementation for our paper, Accurate fatty liver disease diagnosis with a multi-source feature fusion model on the segmented tongue image dataset. We present a non-invasive, reproducible pipeline that fuses features from segmented tongue images with readily obtainable physiological indicators to deliver accurate FLD diagnosis, aiming to support scalable screening in resource-limited settings. The repo provides training and inference code, data preparation, and evaluation protocols to help you reproduce the results and adapt the approach to your own cohorts.
We constructed and publicly released a medium-sized standardized tongue image dataset annotated for FLD. The dataset included the participants' tongue images, basic physiological indicators, and FLD labels. The participants were recruited from a Prospective Cohort of Cancer and Chronic Disease Residents in Fuqing City, which was the first large dynamic prospective cohort study in Fujian Province, China. This cohort study was approved by the Ethics Review Committee of Fujian Medical University (approval numbers [2017-07] and [2020-58]), and written informed consent was obtained from all participants.
The inclusion criteria were as follows: (1) permanent residency in Fuqing City, Fujian Province; (2) age between 35 and 75 years; and (3) completion of both epidemiological data collection and tongue image acquisition between January and December 2021. The exclusion criteria were as follows: (1) incomplete tongue image collection, (2) incomplete epidemiological data, and (3) tongue images classified as noise samples owing to incorrect tongue protrusion posture or motion blur.
We utilized a "TCM Diagnostic Device" equipped with an 8-megapixel camera to capture participants' facial images with protruded tongues. To respect privacy and enhance presentation, we segmented tongues from facial images and released tongue images with a black background. After excluding noise samples, the dataset comprised 5,717 samples, encompassing tongue images of 3,690 Non-FLD participants and 2,027 FLD patients. Additionally, each participant’s data included eight physiological indicators: sex, age, height, waist circumference, hip circumference, weight, systolic blood pressure (SBP), and diastolic blood pressure (DBP).

Our proposed tongue diagnosis framework consisting of four steps: image preprocessing, multi-scale feature extraction, multi-source feature fusion diagnosis, and training or validation.

All procedures followed were in accordance with the ethical standards of the responsible committee on human experimentation (the Ethics Review Committee of Fujian Medical University, China, approval numbers [2017–07] and [2020–58]) and with the Helsinki Declaration of 1975, as revised in 2008 and 2024. Informed consent was obtained from all participants for being included in the study
Python: 3.8.19
pip install -r requirements.txt
unzip ./data/Tongue_Images.zip -d ./data
python main.py
If you find this code or dataset useful in your research, please consider citing us: @inproceedings{gao2024MFF-TDF, title={Accurate fatty liver disease diagnosis with a multi-source feature fusion model on the segmented tongue image dataset}, link={https://github.com/MLDMXM2017/MFF-TDF} }
https://github.com/pytorch/vision/blob/main/torchvision/models/resnet.py https://github.com/ndsclark/MCANet