Inter-vendor performance of deep learning in segmenting acute ischemic lesions on diffusion-weighted imaging: a...

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Radiology
• Neurology

Background:

• Limited evidence exists on the cross-manufacturer applicability of deep learning (DL) for segmenting acute ischemic lesions on diffusion-weighted imaging (DWI).
• Magnetic resonance imaging (MRI) scanner vendor differences can impact the generalizability of DL models.
• Accurate segmentation of ischemic lesions is crucial for timely diagnosis and treatment.

Purpose of the Study:

• To evaluate the inter-vendor operability of DL models for segmenting acute ischemic lesions on DWI.
• To assess the effectiveness of transfer learning in improving the generalizability of DL models across different MRI scanner manufacturers (Siemens and GE).
• To compare the performance of DL models with neuroradiologist performance.

Main Methods:

• Retrospective analysis of DWI data from 6,967 patients with acute ischemic lesions across Siemens and GE MRI scanners.
• Development of two DL models (Model A and Model B) trained on vendor-specific datasets and subsequently fine-tuned using transfer learning.
• Ground-truth segmentation masks created by six neuroradiologists; model performance evaluated using Dice scores against a separate radiologist's segmentations.

Main Results:

• Initial DL models achieved non-inferior performance to radiologists on internal test sets (median Dice scores: 0.858 for Model A, 0.857 for Model B).
• Performance decreased on external test sets, highlighting inter-vendor generalizability challenges.
• Fine-tuned models demonstrated non-inferior performance to radiologists on external test sets (median Dice scores: 0.832 for Model A, 0.846 for Model B), indicating improved cross-vendor applicability.

Conclusions:

• Transfer learning can significantly enhance the inter-vendor operability of DL models for segmenting ischemic lesions on DWI.
• Fine-tuning DL models across datasets improves their clinical applicability and generalizability, addressing limitations posed by different MRI scanner manufacturers.
• This approach holds promise for wider clinical adoption of AI in stroke imaging.