A Neighbor-Sensitive Multi-Modal Flexible Learning Framework for Improved Prostate Tumor Segmentation in Anisotropic...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate prostate tumor segmentation from multi-modal MRI is vital for cancer diagnosis and treatment.
• Current methods struggle with subject-specific modality information and inter-slice data utilization, limiting robustness.
• Anisotropic MR images present challenges due to thick slices and varying modality contributions.

Purpose of the Study:

• To develop a robust and accurate method for segmenting prostate tumors from multi-modal anisotropic MR images.
• To improve the integration of subject-specific information from different MRI modalities.
• To effectively leverage inter-slice information for whole-volume 3D tumor segmentation.

Main Methods:

• Proposed a neighbor-sensitive multi-modal flexible learning network (NesMFle).
• Introduced a Modality-informativeness Flexible Learning (MFLe) module for adaptive multi-modal fusion.
• Utilized a Unet variant with a Sequence Layer and an Activation Mapping Guidance (AMG) module for volumetric segmentation and refinement.
• Employed a random mask strategy during training for improved feature representation.

Main Results:

• NesMFle demonstrated competitive performance on in-house and public (PICAI) multi-modal prostate tumor datasets.
• The method effectively integrates modality-specific information and inter-slice data for accurate tumor delineation.
• Achieved robust and consistent tumor segmentation across neighboring slices.

Conclusions:

• The proposed NesMFle network offers a significant advancement in accurate prostate tumor segmentation from multi-modal anisotropic MR images.
• Flexible learning and inter-slice information utilization are key to overcoming limitations of existing methods.
• NesMFle shows promise for improving clinical diagnosis and treatment planning in prostate cancer.