Evaluating Deep Learning-Based Commercial Software for Detecting Ischemic Lesions on DWI in Stroke Patients.

Area of Science:

• Neurology
• Medical Imaging
• Artificial Intelligence

Background:

• Deep learning advancements enable automated ischemic lesion detection software.
• Real-world performance of AI for diffusion-weighted imaging (DWI) needs further exploration.
• CE-marked AI software (version 1.0) for ischemic lesion detection was evaluated.

Purpose of the Study:

• Evaluate the diagnostic performance of a commercial AI software for detecting ischemic lesions on DWI.
• Assess AI software sensitivity in relation to lesion characteristics.
• Determine AI software's lesion and patient-level sensitivity.

Main Methods:

• Retrospective analysis of 235 patients with confirmed ischemic stroke who underwent DWI.
• Comparison of AI software performance against expert neurologist interpretations (reference standard).
• Analysis of lesion characteristics: size, ADC values, slice coverage, and location.

Main Results:

• The AI software achieved 83.51% lesion-level and 95.31% patient-level sensitivity.
• Undetected lesions were smaller, covered fewer slices, and had higher ADC values.
• Detection rates were not significantly affected by anatomical location, vascular territory, or time from symptom onset.

Conclusions:

• The AI software demonstrates strong overall patient-level sensitivity for ischemic stroke detection.
• Limitations exist in detecting smaller, less conspicuous ischemic lesions.
• Optimization of deep learning algorithms and clinician awareness of AI limitations are crucial for acute stroke care.