🫁 CMRF-Net: Cascade Multi-Resolution Feature Network for Multi-View Chest X-Ray Screening

Official production-grade PyTorch implementation of the Cascade Multi-Resolution Feature Network (CMRF-Net), as published in:

"Diagnostic Accuracy of an Offline CNN Framework Utilizing Multi-View Chest X-Rays for Screening 14 Co-Occurring Communicable and Non-Communicable Diseases" Diagnostics (MDPI) · DOI: 10.3390/diagnostics16010066 · Published: December 2025

Authors: Latika Giri · Pradeep Raj Regmi · Ghanshyam Gurung · Grusha Gurung · Shova Aryal · Sagar Mandal · Samyam Giri · Sahadev Chaulagain · Sandip Acharya · Muhammad Umair

---

🌟 Key Architecture & Features

Component	Description
10-Stage Cascade Encoder-Decoder	SGE → DPEU → HAS → PDE → MSBC → FLR → SUMM → DRB → DRC → RAOH
6 Strategic Skip Connections	Cross-stage spatial links (M2→HAS, M2→SUMM, H3→FLR, S4→MSBC, F6→DRB, C5→DRC)
Dual-Path Expansion Unit (DPEU)	Three parallel branches: spatial, bottleneck, and residual-echo expansion
MultiBranchMixer in SUMM	{1×1, 3×3, 5×5sep, dilated 3×3} parallel branches for small-lesion sensitivity
ASPP Bottleneck in DRC	Atrous Spatial Pyramid Pooling (rates 6/12/18 + global pool)
3-Scale RAOH Detection Heads	256×256, 128×128, 64×64 with GroupNorm (per paper eq.)
Global Triage Head	σ(MLP(GAP(F9))) — 15-class image-level diagnosis (14 pathologies + No Finding)
Custom Vectorized Loss	CIoU (λ=5.0) + Focal obj (λ=1.0) + Focal noobj (λ=0.5) + BCE cls (λ=1.0) + Triage BCE
Offline Deployment	No internet required — tested via PACS integration and standalone mode

---

📁 Repository Structure

ChestXray-Anomaly-Detection/
├── assets/
│   ├── Picture1.png         # Architecture flow diagram
│   └── Picture2.png         # Grad-CAM clinical explainability overlays
├── config/
│   └── config.yaml          # Anchors, loss weights (verified vs. paper), hyperparameters
├── preprocessing/
│   ├── clahe_processor.py   # Percentile contrast stretch + CLAHE (Albumentations)
│   └── annotation_fuser.py  # Enclosing-box removal + WBF (IoU=0.5) + patient-safe splits
└── src/
    ├── model.py             # 10-Stage CMRF-Net — all equations implemented per paper
    ├── dataset.py           # Medical Albumentations augmentations + collation
    ├── loss.py              # CIoU + Focal (obj/noobj separate) + BCE + Triage
    ├── train.py             # Linear warmup (5 ep) + cosine decay, AMP, EMA, early-stop
    ├── eval.py              # mAP, AUC-ROC, Sensitivity, Specificity, NPV, PPV + Grad-CAM
    └── utils.py             # Multi-scale NMS, 14-color bbox drawing, EMA, checkpoint I/O

---

🧠 Stage-by-Stage Architecture

Stage	Module	Output Shape	Key Operations
0	Input Normalization	[B, 3, 512, 512]	MinMax + Z-score (ImageNet stats) + GroupNorm(3)
1	SGE — Shallow Gradient Extractor	[B, 64, 256, 256]	Two 3×3 convs + stride-2 downsampling, α=0.2
2	DPEU — Dual-Path Expansion Unit	[B, 256, 256, 256]	FA / FB / FC parallel branches → concat(320ch) → compress
3	HAS — Hierarchical Aggregation Stack	[B, 256, 256, 256]	4× RepConv blocks + SKIP-1 (M2 concat)
4	PDE — Progressive Depth Encoder	[B, 512, 128, 128]	Stride-2 + DWSepConv × 2 + residual shortcut
5	MSBC — Multi-Scale Bottleneck Cluster	[B, 512, 64, 64]	AdaptiveAvgPool {32,16,8} + SKIP-4 lateral
6	FLR — Feature Lift & Redistribution	[B, 256, 128, 128]	ConvTranspose2d + SKIP-3 (H3 downsampled)
7	SUMM — Secondary Upscale & Mixing	[B, 256, 256, 256]	ConvTranspose2d + SKIP-2 (M2 lateral) + MultiBranchMixer
8	DRB — Downscale Reintegration Block	[B, 256, 128, 128]	Stride-2 + SKIP-5 (F6 concat) → RepConv
9	DRC — Deep Recombination Cluster	[B, 256, 64, 64]	Stride-2 + SKIP-6 (C5 concat) → RepConv → ASPP
10	RAOH Heads × 3 + Global Triage	[B,A,H,W,19] × 3 + [B,15]	GroupNorm detection heads + sigmoid MLP triage

---

🚀 Step-by-Step Usage

1. Install Dependencies

pip install -r requirements.txt

2. Preprocess Images (CLAHE + Resize to 512×512)

python preprocessing/clahe_processor.py \
    --source /path/to/raw_images \
    --dest   /path/to/preprocessed_images \
    --size 512

3. Fuse Annotations (Enclosing Box Filter + WBF + Dataset Splits)

python preprocessing/annotation_fuser.py \
    --csv          /path/to/vinbig2_resized.csv \
    --output       /path/to/fused_annotations.csv \
    --split_dir    /path/to/data_splits \
    --target_size  512 \
    --input_scale  640 \
    --iou_thr      0.5

4. Train

python src/train.py \
    --config        config/config.yaml \
    --train_csv     /path/to/data_splits/train_annotations.csv \
    --val_csv       /path/to/data_splits/val_annotations.csv \
    --img_dir       /path/to/preprocessed_images \
    --checkpoint_dir runs/checkpoints \
    --log_dir        runs/logs

tensorboard --logdir runs/logs

5. Evaluate + Grad-CAM

python src/eval.py \
    --config         config/config.yaml \
    --test_csv       /path/to/data_splits/test_annotations.csv \
    --img_dir        /path/to/preprocessed_images \
    --weights        runs/checkpoints/best_model.pth \
    --gradcam_out    runs/explainability \
    --visualize_count 10

---

👁️ Clinical Explainability with Grad-CAM

CMRF-Net backpropagates gradients through the DRC Stage 9 ASPP bottleneck to generate class-specific visual attention heatmaps, enabling clinically interpretable predictions.

---

📈 Clinical Performance

The following results are directly from the published paper. The model was pretrained on public datasets (VinBig, NIH ChestX-ray14) and fine-tuned on a local dataset from a Nepalese tertiary hospital comprising PA/AP and lateral view chest radiographs annotated by three radiologists for 14 pathologies.

Metric	Value
mAP	0.93
Overall AUC	0.86
Sensitivity	68%
Specificity	99%
Test Set Size	N = 522
Pathology Classes	14
Views Supported	PA, AP, Lateral

Full per-class breakdown and SOTA comparisons are available in the published paper: 10.3390/diagnostics16010066

Deployment Validation

• ✅ Seamless PACS integration confirmed
• ✅ Standalone offline mode validated
• ✅ Robust on poor-quality ER/ICU images
• ✅ Effective on multi-view (PA/AP/Lateral) inputs
• ✅ Colored bounding boxes for co-occurring multi-pathology cases

---

⚙️ Training Hyperparameters

Hyperparameter	Value
Input Resolution	512 × 512
Batch Size	8
Total Epochs	300
Optimizer	AdamW (β₁=0.9, β₂=0.999)
Initial LR	1×10⁻³
Min LR	1×10⁻⁶
LR Schedule	Linear warmup (5 ep) + Cosine annealing
Weight Decay	5×10⁻⁴ (conv weights only)
Gradient Clip	θ = 10.0
EMA Decay	0.9999
Early Stopping Patience	20 epochs
Mixed Precision	AMP (FP16/FP32)

Loss Weights

Weight	Symbol	Value	Description
Localization (CIoU)	λ_box	5.0	Emphasis on precise bounding box regression
Objectness — positive	λ_obj	1.0	Focal loss on cells containing a pathology
Objectness — background	λ_noobj	0.5	Background cell suppression penalty
Classification (BCE)	λ_cls	1.0	Multi-label pathology class loss
Triage head	λ_triage	1.0	Global 15-class image-level classification
Focal loss alpha	α_focal	0.25	Foreground/background balance factor
Focal loss gamma	γ_focal	2.0	Focusing exponent (hard-example weighting)
Ignore IoU threshold	τ_ignore	0.5	Suppresses gradient on ambiguous overlaps

---

📄 Citation

@Article{diagnostics16010066,
  AUTHOR   = {Giri, Latika and Regmi, Pradeep Raj and Gurung, Ghanshyam and
              Gurung, Grusha and Aryal, Shova and Mandal, Sagar and
              Giri, Samyam and Chaulagain, Sahadev and Acharya, Sandip and
              Umair, Muhammad},
  TITLE    = {Diagnostic Accuracy of an Offline CNN Framework Utilizing
              Multi-View Chest X-Rays for Screening 14 Co-Occurring
              Communicable and Non-Communicable Diseases},
  JOURNAL  = {Diagnostics},
  VOLUME   = {16},
  YEAR     = {2025},
  NUMBER   = {1},
  PAGES    = {66},
  DOI      = {10.3390/diagnostics16010066},
  URL      = {https://www.mdpi.com/2075-4418/16/1/66}
}

---

🛡️ License

This project is licensed under the MIT License. The paper is open access under CC-BY 4.0.