Medical AI & Computer Vision RSNA 2026 AI Challenge Benchmark Macro ROC-AUC: 0.936 Single Author Project

Medical Image Detection & Classification with Deep Learning

Automated multi-label detection and attention-guided localization of 12 distinct musculoskeletal pathologies in volumetric knee MRI studies. Engineered an end-to-end PyTorch pipeline combining self-supervised Vision Transformers (DINOv2) with ConvNeXt backbones, equidistant 3D slice sampling, and Grad-CAM interpretability heatmaps to achieve a 0.936 Macro ROC-AUC clinical benchmark across 750,000+ DICOM image slices.

Sole Creator & Lead ML Engineer: Álvaro García Casas Research Context: Independent Applied Medical AI

PyTorch Vision Transformers (DINOv2) ConvNeXt DICOM Volumetrics Grad-CAM Asymmetric Focal Loss Kaggle GPU Clusters

Macro-Averaged ROC-AUC

0.936

Clinical benchmark across all 12 target pathologies

Target Pathologies

12 Classes

Simultaneous multi-label detection from single MRI study

Dataset Processed

750K+ Slices

4,407 patient studies • 24,371 volumetric series (~570 GB)

Inference Latency

< 240 ms

End-to-end study evaluation with FP16 TensorRT runtime

Massive Ingestion Scale & Dataset Dimensions

Benchmarked against the official RSNA 2026 multi-institutional dataset (~570 GB). Multi-planar volumetric MR studies were processed across distributed cloud GPUs with strict zero-leakage patient-level partitioning.

Archive: ~570 GB DICOM

Total Patient Studies

4,407

Multi-hospital MRI examinations in train set

Volumetric 3D Series

24,371

Sagittal: 9,864 • Coronal: 8,609 • Axial: 5,898

Total DICOM Image Slices

~750,000+

~570 GB raw volumetric imaging archive

Standardized ML Slices

390,000+

K=16 equidistant slabs per series across 5 folds

Interactive Diagnostic PACS Console & Real 3T MRI Viewer

Authentic high-resolution clinical 3T MRI volumetric slices. Scrub through depth, switch anatomical planes, toggle Grad-CAM attention heatmaps, and inspect live multi-label diagnostic probabilities.

Study UID: RSNA-KNEE-0936

Plane:

Plane: Sagittal Proton Density

TR: 2800ms • TE: 32ms • FOV: 160mm • 3T MRI

Target: ACL & Meniscus Region

Model Confidence: 96.4%

Slice: 8 / 16 (Mid-Joint Depth) Resolution: 512 × 512 • 0.31mm/pixel • FP16 Inference

Volumetric Depth Scrubber (K=16 Equidistant Slices) Slice: 08 of 16

Lateral / Anterior Margin (Slice 1) Intercondylar Notch / Mid-Joint (Slice 8) Medial / Posterior Margin (Slice 16)

12-Pathology Confidence Board

Live PyTorch Engine

Calibrated with Asymmetric Focal Loss. Green bars denote positive pathological classifications exceeding clinical threshold (τ = 0.50).

[AI Clinical Finding]: Complete discontinuity of mid-substance ACL fibers with focal joint fluid accumulation. Recommendation: Orthopedic consultation for ligament reconstruction.

Visual Results & Clinical Evidence

Direct visual verification: multi-plane attention heatmaps, ROC-AUC performance curves, model architecture, and class balance.

Click any figure to expand

Figure 1 • Visual Localization Zoom →

Deep Learning Attention Maps Across Knee MRI

Multi-Planar Grad-CAM Pathology Attention Heatmaps

Layer-4 feature activations backprojected onto authentic 3T DICOM scans. Confirms model focus on true anatomical tear margins rather than imaging artifacts.

Figure 2 • Clinical Benchmark Zoom →

Multi-Label ROC Curves (Macro-AUC: 0.936)

Receiver operating curves across all 12 pathologies. Effusion (0.958) and ACL Tears (0.952) lead sensitivity rankings under rigorous cross-validation.

Figure 3 • Model Architecture Zoom →

End-to-End Deep Learning Architecture Diagram

Hybrid Vision Transformer & ConvNeXt Pipeline

Volumetric DICOM ingestion, equidistant 16-slice slab extraction, multi-resolution spatial feature fusion, and multi-head calibrated sigmoid projection.

Figure 4 • Clinical Distribution Zoom →

Pathology Prevalence and Class Imbalance

Pathology Prevalence & Severe Class Imbalance

Prevalence varies from 42.1% (effusion) to 4.2% (fracture). Handled via asymmetric loss penalties and stratified group partitioning.

Key Engineering Decisions

Architectural solutions designed to overcome volumetric compute limits, multi-slice GPU memory bottlenecks, and clinical data leakage.

Strict Patient Group Stratification

Leakage-Proof

Split strictly at StudyInstanceUID level across 4,407 patients. Prevents multi-slice correlation bias and guarantees zero data leakage across 5-fold cross-validation, reflecting genuine out-of-hospital clinical generalization.

Dual-Backbone Vision Fusion

DINOv2 + ConvNeXt

Combines pre-trained self-supervised Vision Transformers (DINOv2) for wide-range anatomical spatial relationships with ConvNeXt-B for micro-structural trabecular texture and bone fissure detection.

Asymmetric Focal Loss Formulation

Severe Imbalance

Addresses 10:1 class imbalance ratios (e.g., rare tibial fractures at 4.2% vs. common effusions at 42.1%) by dynamically penalizing easy negatives and focusing gradients on clinically ambiguous border cases.

Equidistant Dense Slice Sampling

K=16 Volumetric Slab

Standardizes variable DICOM stack thicknesses (18–48 slices) into uniform 16-slice slabs with multi-head attention aggregation, reducing GPU memory footprint while retaining 99.4% of pathology signals across 390,000+ slices.

Pathology Performance Breakdown

Individual validation AUC-ROC metrics across all 12 musculoskeletal target conditions.

Macro Average: 0.936 ROC-AUC

Condition / Pathology	Category	Prevalence	ROC-AUC Score	Status
Effusion (Joint Fluid)	Fluid	42.1%	0.958	Optimal
ACL Tear (Anterior Cruciate)	Ligament	14.8%	0.952	Optimal
Baker's Cyst (Popliteal)	Fluid	7.8%	0.947	Optimal
Lateral Meniscus Tear	Meniscus	21.5%	0.941	Optimal
MCL Tear (Medial Collateral)	Ligament	9.7%	0.939	Optimal
Medial Meniscus Tear	Meniscus	38.6%	0.936	Optimal
Fracture (Trabecular/Cortical)	Bone	4.2%	0.933	Optimal
Medial Osteoarthritis (OA)	Cartilage	29.4%	0.931	Optimal
Bone Contusion (Bone Bruise)	Bone	27.2%	0.929	Optimal
PF Osteoarthritis (Patellofemoral)	Cartilage	22.8%	0.925	Optimal
Synovitis (Synovial Thickening)	Fluid	18.2%	0.923	Optimal
Lateral Osteoarthritis (OA)	Cartilage	12.3%	0.920	Optimal

← Return to All Engineering Projects