Simultaneous image reconstruction and lesion segmentation in accelerated MRI using multitasking learning.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Magnetic Resonance Imaging (MRI) is crucial but limited by long scan times, causing patient discomfort and image artifacts.
• Manual lesion segmentation is time-consuming, and automated methods struggle with low-quality, accelerated imaging.

Purpose of the Study:

• To develop a multitask learning method (RecSeg) for simultaneous MRI reconstruction and lesion segmentation.
• To evaluate if combined reconstruction and segmentation tasks benefit from a unified model.

Main Methods:

• Proposed a novel multitask learning framework, RecSeg, integrating image reconstruction and lesion segmentation.
• Utilized two connected U-Nets for liver and renal image reconstruction and segmentation.
• Validated the model on MR k-space data with acceleration factors of 2×, 4×, and 6×, using data from 50 healthy subjects and 100 hepatocellular carcinoma patients.

Main Results:

• RecSeg achieved superior image reconstruction quality, evidenced by higher PSNR and SSIM compared to KSVD and single U-net at 6× acceleration.
• Demonstrated improved lesion segmentation accuracy with the highest Dice score for RecSeg.
• Results indicate significant performance gains (p < 0.05) in both reconstruction and segmentation tasks.

Conclusions:

• The multitask learning approach effectively addresses challenges in accelerated MRI.
• Simultaneous image reconstruction and lesion segmentation using RecSeg enhances overall performance.
• This method holds promise for faster and more accurate MRI analysis in clinical settings.