FracAtlas Bone Fracture Classification & Segmentation

End-to-End CAD Radiographic Framework Combining DenseNet-121, ResNet-34 U-Net, and Meta's SAM for Zero-Shot Boundary Delineation.

Overview

Accurate radiological interpretation of bone fractures requires both rapid patient-level screening and fine-grained spatial localization. Subtle hairline fissures, complex peri-articular dislocations, and anatomical overlaps across diverse projections (frontal, lateral, oblique) frequently challenge conventional computer-aided diagnosis (CAD) pipelines.

This project delivers an end-to-end multi-task deep learning framework (FracAtlas) designed for automated bone fracture detection and pixel-level lesion segmentation from digital radiographs. The pipeline integrates a decoupled architecture: global binary classification via DenseNet-121, dense semantic segmentation via a ResNet-34 U-Net, and zero-shot foundation boundary delineation using Meta’s Segment Anything Model (SAM) guided by geometric bounding-box prompts.

Architectural Pipeline

  • Data Engineering & Stream Sanitization: Ingests raw radiography streams, handling truncated or corrupt JPEG encodings via low-level PIL/C-library handlers (ImageFile.LOAD_TRUNCATED_IMAGES), paired with dynamic Albumentations transforms.
  • Global Classification Head (DenseNet-121): Predicts patient-level fracture probability with dense feature reuse and direct gradient backpropagation paths to mitigate vanishing gradients.
  • Semantic Localization Head (ResNet34-UNet): Decodes fine-grained pixel masks trained via a composite BCEWithLogitsLoss and DiceLoss objective to counter extreme background-to-lesion class imbalance.
  • Zero-Shot Boundary Prompting (Meta SAM ViT-B): Integrates vision foundation models conditioned on COCO bounding-box coordinates to isolate complex boundary geometries and non-displaced diaphyseal shaft fractures.
  • Decoupled Inference Engine: Executes conditional segmentation gating and visualizes global classification confidence alongside side-by-side ground truth and predicted spatial masks.

Technical Highlights & Methodological Insights

1. Multi-Task Decoupling & Sensitivity Calibration

  • Calibrated Decision Boundaries: To maximize clinical recall in emergency room triage, the segmentation inference head utilizes a lowered probability threshold ($0.20 \le \tau \le 0.30$), prioritizing lesion sensitivity over conservative precision.
  • Diagnostic Discordance Handling: Empirical evaluations reveal cases where global classifiers report low fracture confidence ($P < 0.05$) while the spatial segmentation head resolves exact morphological boundaries, demonstrating the necessity of decoupled, multi-tier CAD architectures.

2. Anatomical Generalization

  • High-density joint and peri-articular structures (such as the distal radius, wrist, and ankle) achieve superior segmentation precision ($\text{IoU} > 0.90$, maximum probability $\ge 0.98$).
  • For hairline fractures lacking obvious cortical disruption, explicit geometric prompt engineering via SAM ViT-B recovers crisp topological contours where standard convolutional backbones experience boundary blurring.

Dataset & Annotation Schema

Evaluated on the benchmark FracAtlas Dataset (Nature Scientific Data):

  • Payload: 4,083 clinical radiographs across hand, leg, hip, and shoulder anatomies.
  • Labels & Targets: 717 confirmed fractured cases and 3,366 healthy scans paired with dense COCO polygon masks and bounding boxes.
  • Projections: Multi-view evaluations spanning frontal (AP/PA), lateral, and oblique views.

Technical Stack

  • Deep Learning Runtime: Python 3.8+, PyTorch, Torchvision
  • Segmentation Architectures: segmentation-models-pytorch, Meta Segment Anything Model (segment-anything)
  • Computer Vision & Augmentations: OpenCV, Albumentations, PIL, PyCOCOTools
  • Evaluation & Analytics: Scikit-Learn (IoU, Dice, Precision, Recall, ROC-AUC), NumPy, Pandas