AI-Assisted Ultrasound Analysis

Fatty Liver Detection System

AI-powered classification of liver ultrasound images using Siamese Neural Networks (ResNet-50)

B-Mode Ultrasound 5 Steatosis Grades ResNet-50 Backbone Real-Time Inference

Image Classification

Classify liver ultrasound images into 5 categories

Deep Learning Model

Siamese Neural Network with ResNet-50 encoder

5-Class Prediction

Normal, Grade I, Grade II, Grade III, CLD

Fast & Easy to Use

Upload an image and get instant prediction results

Diagnostic Scans

--

Total scans archived in cloud

AI Engine

ResNet-50

Siamese Deep Feature Embedder

Cloud Database

Supabase

Active live synchronization

Inference Speed

~180ms

Real-time CPU execution

Start Ultrasound Steatosis Examination

Upload a B-mode ultrasound DICOM/image or test with verified reference scans in the dedicated clinical workstation. Instant 5-stage staging, acoustic attenuation profiling, and ISO A4 PDF report generation.

Patient & Examination Information (Included in Downloadable A4 Report)

Editable • Auto-populated from clinician session

Patient ID / MRN

Patient Full Name

Age / Gender

Examination Date

Clinical Indication / Reason for Exam

Upload Ultrasound Image

Drag and drop an ultrasound image here

or

Supported formats: JPG, JPEG, PNG (Max 16MB)

image.jpg

-- × -- • -- KB

Sample Ultrasound Image

Normal Control Scan

Homogeneous parenchyma, normal liver-to-kidney contrast, preserved portal vein wall echogenicity, and sharp diaphragm contour.

Predicted Class Probabilities

Normal NORMAL

Confidence: --% Inference Latency: --s

Healthy liver with no steatosis. Liver echogenicity is normal.

PREDICTED

Normal

< 5% Steatosis

0.0%

PREDICTED

Grade I

5% - 35% Steatosis

0.0%

PREDICTED

Grade II

35% - 65% Steatosis

0.0%

PREDICTED

Grade III

> 65% Steatosis

0.0%

PREDICTED

CLD

Chronic Liver Disease

0.0%

Ultrasound Analytics & Clinical Detail Graphs

Biomarker Engine Active

Quantitative acoustic sound decay, multi-class probability metrics, parenchymal echogenicity, and steatosis radar

Mean Echogenicity

88.5 / 255

Attenuation Gradient

-0.31 dB/cm

Speckle Noise σ²

±26.4

Diagnostic Status

Normal Control

Medical Research Disclaimer: This application is intended for research and educational purposes. Its predictions are AI-generated estimates and are not a substitute for professional medical assessment, radiologist review, or official clinical diagnosis. Always consult a qualified medical professional regarding liver health concerns.

Prediction History

Record ID	Date & Time	Filename	Mode	Prediction	Confidence	Severity	Action

No prediction history recorded yet. Run an analysis from the dashboard.

Siamese Neural Network Architecture

Twin ResNet-50 feature extractor with contrastive self-supervised pre-training and downstream 5-class linear classifier.

Supported Steatosis Categories

Normal < 5% fat, normal parenchymal echogenicity

Grade I 5% - 35% fat, mild diffuse hyperechogenicity

Grade II 35% - 65% fat, impaired portal vessel definition

Grade III > 65% fat, marked acoustic attenuation

CLD Chronic Liver Disease, fibrotic parenchymal changes

Dataset: Evaluated on the IJCARS clinical B-mode liver ultrasound benchmark (55 patients, 550 raw scans) with self-supervised contrastive learning and spatial augmentations.

About the Fatty Liver Detection System

Non-Alcoholic Fatty Liver Disease (NAFLD) affects up to 25% of the global adult population and is strongly associated with metabolic syndrome, type 2 diabetes, and cardiovascular conditions. Early and accurate detection of hepatic steatosis is essential to halt or reverse progression toward non-alcoholic steatohepatitis (NASH), cirrhosis, or hepatocellular carcinoma.

While liver biopsy is the historic reference standard, it is invasive, costly, and associated with patient discomfort and complications. Ultrasound B-mode imaging is widely accessible and non-invasive, but visual qualitative evaluation suffers from inter-observer and intra-observer variability.

This clinical AI system implements a Siamese Neural Network with a ResNet-50 backbone trained via self-supervised contrastive learning. By extracting high-dimensional latent representations of liver parenchyma and ultrasound attenuation patterns, the model provides objective, reproducible assessments across all five clinical steatosis categories.

Clinical Usage Protocol

Ensure the ultrasound image captures liver parenchyma adjacent to the right kidney or diaphragm.
Verify the scan is in standard B-mode format (JPG, PNG, or BMP).
Upload the scan or select one of the clinical reference samples.
Review the multi-class probability distribution and confidence score.
Export or print the report for medical records and radiologist corroboration.

Clinical Profile & Hospital Header

Included on the official downloadable A4 medical reports

Healthcare Facility / Hospital

Institution name on printed clinical documentation

Clinical Department

Medical division or imaging center

Attending Radiologist / Clinician

Name appearing on report electronic signature

Medical License / ID

Physician accreditation or PACS terminal identifier

AI Inference & Diagnostic Staging

Configure deep learning behavior and reporting defaults

Default Classification Mode

Preset active mode when loading the workstation

Confidence Alert Threshold

Flag predictions below this threshold as uncertain

75%

Acoustic Biomarker Profiling

Calculate sound decay and hepatorenal contrast ratios

Auto-Open A4 Report Preview

Display the printable A4 sheet modal right after inference

Backend API Server

Inference engine endpoint and connectivity diagnostics

API Base URL

Where ultrasound images are dispatched for model inference

Quick Presets

Switch between deployed cloud and local development

Connection Health

Not tested in this session

Cloud Database & Appearance

Supabase synchronization and visual theme preferences

Supabase Cloud Sync

Automatically archive patient predictions in Supabase table

Supabase Credentials

Connected to production instance

Interface Theme

Switch between clean medical light and dark PACS modes