Detecting lumbar lesions in ^99m Tc-MDP SPECT by deep learning: Comparison with physicians.

Area of Science:

• Nuclear Medicine
• Radiology
• Artificial Intelligence in Medical Imaging

Background:

• 99m Tc-MDP SPECT is used for diagnosing pediatric lumbar stress, a cause of low back pain.
• Detection of small lesions is challenging due to SPECT image quality and interreader variability.
• Deep learning offers potential for improved lesion detection.

Purpose of the Study:

• Develop a deep convolutional neural network (CNN) for detecting lumbar lesions in 99m Tc-MDP scans.
• Compare the CNN's performance against physicians using a localization receiver operating characteristic (LROC) study.
• Improve diagnostic accuracy for pediatric lumbar stress injuries.

Main Methods:

• Developed an automated method to insert artificial lesions into lesion-absent 99m Tc-MDP scans to create realistic lesion-present images.
• Trained a deep CNN using 2.5D views from these generated images.
• Evaluated the CNN's performance using cross-validation on a test dataset, comparing it with five physicians in an LROC study.

Main Results:

• The deep learning (DL) system achieved an area under the LROC curve (AUC_LROC) of 0.830, slightly outperforming the overall physician performance (AUC_LROC = 0.785).
• The DL system's performance was comparable or superior to that of individual physicians in localizing small 99m Tc-MDP positive lumbar lesions.
• The DL system demonstrated potential for reducing interreader variability in diagnosing pediatric lumbar stress.

Conclusions:

• The developed DL system shows comparable or superior performance to physicians in localizing small 99m Tc-MDP positive lumbar lesions.
• This AI approach may enhance the diagnostic accuracy and consistency of 99m Tc-MDP SPECT for pediatric low back pain evaluation.
• Further validation is warranted to integrate DL into clinical practice for pediatric lumbar stress diagnosis.