{% extends "base.html" %} {% block title %}About — ICH Screening{% endblock %} {% block content %}

About This System

AI-Assisted CT-Based Intracranial Hemorrhage Detection with Explainability and Clinical Reporting

System Overview

This is an AI-assisted screening tool designed to detect intracranial hemorrhage (ICH) from CT brain scans. It combines deep learning with visual explainability, confidence calibration, and structured clinical reporting to support — not replace — medical decision-making.

1
CT Brain Image Input
→
2
Preprocessing & CT Windowing
→
3
2.5D Detection (EfficientNet-B4)
→
4
Grad-CAM Explainability
→
5
Confidence Calibration
→
6
Clinical Report

Model Architecture

ArchitectureEfficientNet-B4 (timm)
Input Representation2.5D (prev/center/next)
Channels9 (3 CT windows × 3 slices)
Outputs6 heads (any + 5 subtypes)
Inference Strategy5-fold ensemble (logit averaging)

CT Preprocessing

Brain WindowWC=40, WW=80
Subdural WindowWC=75, WW=215
Soft Tissue WindowWC=40, WW=380
Channels3 (one per window)
FormatDICOM → HU → windowed RGB

Calibration

Method{{ calib.get('method', calib.get('best_method', 'N/A')) }}
{% if calib %}
Temperature{{ '%.4f'|format(calib.temperature) }}
Threshold{{ '%.4f'|format(calib.calibrated_threshold) }}
{% endif %}
ECE (Raw → Calibrated){{ '%.4f'|format(calib.get('raw_ece', 0.0)) }} → {{ '%.4f'|format(calib.get('cal_ece', 0.0)) }}
Bands HIGH (≥{{ '%.2f'|format(calib.get('high_threshold', 0.7)) }}) · MEDIUM ({{ '%.2f'|format(calib.get('low_threshold', 0.3)) }}–{{ '%.2f'|format(calib.get('high_threshold', 0.7)) }}) · LOW (<{{ '%.2f'|format(calib.get('low_threshold', 0.3)) }})

Explainability

MethodGrad-CAM
Target LayerLast convolutional block
OutputHeatmap overlay on input
PurposeVisual evidence for review

Confidence-Aware Triage System

Instead of a simple binary output, the system incorporates prediction confidence into a three-band triage workflow:

HIGH ≥ {{ '%.2f'|format(calib.get('high_threshold', 0.7)) }} calibrated probability

If positive: Urgent radiologist review recommended

If negative: Standard workflow — no urgent action

MEDIUM {{ '%.2f'|format(calib.get('low_threshold', 0.3)) }} – {{ '%.2f'|format(calib.get('high_threshold', 0.7)) }}

If positive: Prioritised radiologist review recommended

If negative: Standard workflow — manual review if clinically indicated

LOW < {{ '%.2f'|format(calib.get('low_threshold', 0.3)) }}

If positive: Radiologist review recommended — low confidence

If negative: Manual review recommended — model uncertainty high

Dataset

SourceRSNA Intracranial Hemorrhage Detection
ModalityCT brain (axial slices)
FormatDICOM
TaskAny-hemorrhage screening + subtype-aware outputs

Ethical Considerations & Limitations

This System Is:

  • A screening and decision-support tool
  • Designed to assist, not replace, medical professionals
  • Transparent via Grad-CAM visual evidence
  • Calibrated for reliable confidence scores
  • Built on publicly available, ethically sourced data

This System Is NOT:

  • A diagnostic device or medical diagnosis tool
  • A replacement for qualified radiologist review
  • Cleared for standalone clinical deployment
  • A substitute for clinical subtype confirmation
  • Validated for real-time hospital use
Important Disclaimer: This system is produced by an AI-assisted screening tool and does NOT constitute a medical diagnosis. All screening findings must be reviewed and confirmed by a qualified, licensed medical professional before any clinical decision is made. The system is intended solely as a decision-support aid in a screening workflow and is not cleared for standalone diagnostic use.

Technology Stack

Python PyTorch EfficientNet-B4 timm 2.5D Context 5-Fold Ensemble Isotonic Calibration OpenCV NumPy Pandas Matplotlib Grad-CAM Flask pydicom scikit-learn
{% endblock %}