A deep learning algorithm for detecting lytic bone lesions of multiple myeloma on CT

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Medicine
• Oncology

Background:

• Whole-body low-dose CT is standard for assessing multiple myeloma bone destruction.
• Manual interpretation of CT scans for small lytic lesions is time-consuming.
• Deep learning (DL) offers potential to enhance CT scan utility in myeloma imaging.

Purpose of the Study:

• To develop a DL algorithm for detecting multiple myeloma lytic bone lesions.
• To evaluate the diagnostic performance of the developed DL model.

Main Methods:

• A two-step DL model was created, involving bone segmentation (Unet) and lesion detection (YOLO).
• The model was trained and tested on axial slices from low-dose CT scans of patients with plasma cell dyscrasias.
• Diagnostic performance was assessed using the area under the receiver operating characteristic curve (AUROC).

Main Results:

• The DL model achieved a sensitivity of 91.6% and a specificity of 84.6%.
• The lesion detection AUROC was 90.4%.
• The study analyzed 2193 slices from 40 patients, identifying 5640 lytic lesions.

Conclusions:

• A high-performing DL model for detecting multiple myeloma lytic bone lesions on low-dose CT has been developed.
• The model demonstrates significant potential for improving diagnostic accuracy and efficiency.
• Further external validation is necessary before clinical implementation.