CheXVision-DenseNet

CheXVision — Deep Learning & Big Data university project. 14-class chest X-ray pathology detection + binary normal/abnormal classification on the NIH Chest X-ray14 dataset (112,120 images).

Project Resources

• GitHub repository
• Presentation deck
• Live demo
• Dataset

Architecture

Fine-Tuning Strategy

Training Pipeline

Training Metrics

• Best validation macro AUC-ROC: 0.8459
• Best validation binary AUC-ROC: 0.7867
• Best validation binary F1: 0.6736
• Best checkpoint epoch: 18

Per-Class AUC-ROC at Best Epoch

Pathology	AUC-ROC	Visual
Atelectasis	0.8334	████████░░
Cardiomegaly	0.9010	█████████░
Effusion	0.8873	█████████░
Infiltration	0.7133	███████░░░
Mass	0.8756	█████████░
Nodule	0.8084	████████░░
Pneumonia	0.7397	███████░░░
Pneumothorax	0.8705	█████████░
Consolidation	0.8063	████████░░
Edema	0.9255	█████████░
Emphysema	0.9107	█████████░
Fibrosis	0.8085	████████░░
Pleural_Thickening	0.8377	████████░░
Hernia	0.9242	█████████░

Training Configuration

• Repository: arudaev/chexvision-densenet
• Dataset: arudaev/chest-xray-14-320 · revision 44443e6ee968b3c6094b63f14a27698c40b50680
• Architecture: DenseNet-121 transfer learning with a shared feature layer and dual classification heads.
• Platform: Kaggle GPU kernel (NVIDIA T4 / P100)
• Batch size: 24 × grad_accum 4 = effective batch 96
• AMP (fp16): enabled
• CLAHE preprocessing: enabled
• Label smoothing: 0.1
• Optimizer: AdamW · Scheduler: CosineAnnealingLR
• Epochs configured: 60 · Early stop patience: 15

Intended Use

This model is intended for research and educational work on automated chest X-ray pathology detection. It outputs two predictions per image:

1. Multi-label scores — independent sigmoid probability for each of 14 NIH pathologies
2. Binary score — sigmoid probability of any abnormality (Normal vs. Abnormal)

Limitations

• Not validated for clinical use. Predictions must not substitute professional medical judgment.
• Trained on NIH Chest X-ray14, which contains noisy radiologist annotations (patient-level labels, not lesion-level).
• Performance degrades on images from equipment, patient populations, or preprocessing pipelines that differ from the NIH training distribution.
• Reported AUC metrics are on the validation split, not the held-out test set.

CheXNet Benchmark Context

CheXNet (Rajpurkar et al., 2017) — the seminal paper establishing DenseNet-121 for chest X-ray classification — reported 0.841 macro AUC-ROC on a comparable split of this dataset. CheXVision-DenseNet reaches 0.8459 macro AUC-ROC — slightly exceeding this benchmark on the validation split, trained at 320×320 resolution with an added binary head under a fixed Kaggle GPU budget. See the CheXVision demo for live inference, or the presentation deck for the project walkthrough.

Citation

@misc{chexvision2026,
  title={CheXVision: Dual-Task Chest X-ray Classification with Custom CNN and DenseNet-121},
  author={BIG D(ATA) Team},
  year={2026},
  howpublished={\url{https://huggingface.co/arudaev/chexvision-densenet}}
}