A Great Place to Diagnose Thoracic Diseases
ChestVision-AI is a high-performance clinical computer vision system developed to detect 14 distinct thoracic pathologies simultaneously from chest radiographs using a dual-branch ConvNeXt-Autoencoder fusion architecture.
Dual-Branch Fusion
Unsupervised Autoencoder extracts 512-D spatial vectors, seamlessly fused with ConvNeXt-Base 1024-D hierarchical semantic features.
14 NIH Pathologies
Multi-label classification with per-disease Youden's J-statistic dynamic thresholds rather than arbitrary 0.5 cutoffs.
Zero-Leakage Cohort
Strict patient-isolated partitioning across 24,000+ radiographs guarantees clinical independence between train, validation, and test splits.
Leading Thoracic Diagnostic Studio
Real-time visual explainability with Grad-CAM heatmaps, Autoencoder latent difference maps, DICOM windowing, and per-disease calibrated probability radar.
"ChestVision-AI's dual-branch architecture detected subtle Cardiomegaly and bilateral Pleural Effusion with pinpoint anatomical localization."
Select a Clinical Benchmark Case from NIH Cohort
6 authentic NIH benchmark cases available with instant multi-label diagnostic inference.
64 Y / Male • PA View
AI Diagnostic Impression & Findings
Marked cardiomegaly with bilateral pleural effusions (predominantly right-sided) and dependent bibasilar congestion.
Dual-Branch Hybrid Fusion Architecture
Unsupervised Convolutional Autoencoder structural embeddings combined with hierarchical ConvNeXt-Base semantic representations.
ConvNeXt-Base Branch
Extracts 1024-dim deep hierarchical semantic feature maps from radiographs.
1024-D VectorsAutoencoder Branch
Unsupervised 512-dim latent spatial vector capturing lung-field structural anomalies.
512-D LatentFeature Fusion Layer
Latent concatenation into 1536-dim vector with BatchNorm and 0.40 Dropout.
1536-D Fusion14-Head Calibrated Sigmoids
Independent multi-label classification calibrated by Youden’s J-statistic.
14 PathologiesAblation Study (NIH Benchmark)
| Architecture | ROC-AUC | F1-Score | Specificity | Latency |
|---|---|---|---|---|
| ConvNeXt + AE (Ours) | 0.874 | 0.768 | 91.4% | 42.8 ms |
| ConvNeXt-Base Alone | 0.838 | 0.712 | 88.2% | 36.2 ms |
| DenseNet-121 (CheXNet) | 0.815 | 0.680 | 85.9% | 22.4 ms |
| ResNet-50 Baseline | 0.792 | 0.654 | 83.1% | 19.8 ms |
| Autoencoder Alone | 0.710 | 0.540 | 78.4% | 14.1 ms |
Multi-Disease ROC-AUC Curves
Our Clinical & AI Research Team
Expert pulmonologists, medical imaging specialists, and deep learning researchers dedicated to advancing thoracic diagnosis.