Real-world application of a 3D deep learning model for detecting and localizing cerebral microbleeds

Background

Detection and localization of cerebral microbleeds (CMBs) is crucial for disease diagnosis and treatment planning. However, CMB detection is labor-intensive, time-consuming, and challenging owing to its visual similarity to mimics. This study aimed to validate the performance of a three-dimensional (3D) deep learning model that not only detects CMBs but also identifies their anatomic location in real-world settings.

Methods

A total of 21 patients with 116 CMBs and 12 without CMBs were visited in the neurosurgery outpatient department between January 2023 and October 2023. Three readers, including a board-certified neuroradiologist (reader 1), a resident in radiology (reader 2), and a neurosurgeon (reader 3) independently reviewed SWIs of 33 patients to detect CMBs and categorized their locations into lobar, deep, and infratentorial regions without any AI assistance. After a one-month washout period, the same datasets were redistributed randomly, and readers reviewed them again with the assistance of the 3D deep learning model. A comparison of the diagnostic performance between readers with and without AI assistance was performed.

Results

All readers with an AI assistant (reader 1:0.991 [0.930–0.999], reader 2:0.922 [0.881–0.905], and reader 3:0.966 [0.928–0.984]) tended to have higher sensitivity per lesion than readers only (reader 1:0.905 [0.849–0.942], reader 2:0.621 [0.541–0.694], and reader 3:0.871 [0.759–0.935], p = 0.132, 0.017, and 0.227, respectively). In particular, radiology residents (reader 2) showed a statistically significant increase in sensitivity per lesion when using AI. There was no statistically significant difference in the number of FPs per patient for all readers with AI assistant (reader 1: 0.394 [0.152–1.021], reader 2: 0.727 [0.334–1.582], reader 3: 0.182 [0.077–0.429]) and reader only (reader 1: 0.364 [0.159–0.831], reader 2: 0.576 [0.240–1.382], reader 3: 0.121 [0.038–0.383], p = 0.853, 0.251, and 0.157, respectively). Our model accurately categorized the anatomical location of all CMBs.

Conclusions

Our model demonstrated promising potential for the detection and anatomical localization of CMBs, although further research with a larger and more diverse population is necessary to establish clinical utility in real-world settings.