An explainable deep-learning algorithm for the detection of acute intracranial haemorrhage from small datasets

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Deep learning algorithms offer potential for automated medical diagnoses but often function as "black boxes."
• High-performance deep learning necessitates large, high-quality training datasets, which can be a limitation in clinical settings.
• Explainability and data efficiency are critical challenges for integrating AI into clinical decision-making.

Purpose of the Study:

• To develop an understandable deep-learning system for detecting and classifying acute intracranial hemorrhage (ICH) and its subtypes.
• To improve the explainability of deep learning models in medical image analysis.
• To demonstrate high performance using a relatively small dataset, mimicking radiologist workflows.

Main Methods:

• Development of a deep-learning system utilizing unenhanced head computed-tomography (CT) scans.
• Training the algorithm on a dataset of 904 cases for ICH detection and subtype classification.
• Incorporation of an attention map and prediction basis for enhanced explainability.
• Implementation of an iterative process designed to emulate the diagnostic workflow of radiologists.

Main Results:

• The system achieved performance comparable to expert radiologists on independent test datasets.
• Sensitivity of 98% and specificity of 95% on a 200-case test dataset.
• Sensitivity of 92% and specificity of 95% on a 196-case test dataset.
• Demonstrated effectiveness with a limited training dataset (904 cases).

Conclusions:

• The developed deep-learning system provides an explainable and accurate method for diagnosing acute intracranial hemorrhage (ICH) from head CT scans.
• The system's performance, explainability features, and data efficiency suggest a viable approach for clinical integration.
• This methodology can facilitate the development and adoption of deep learning systems across various clinical applications, accelerating AI integration into medical practice.