MMO-Net (Multi-Magnification Organ Network): A use case for Organ Identification using Multiple Magnifications in...

Area of Science:

• Digital Pathology
• Computational Toxicology
• Artificial Intelligence in Pathology

Background:

• Accurate organ identification in histology images is crucial for toxicological digital pathology.
• Previous automated methods using single magnifications struggle with small or contiguous organs on whole slide images (WSIs).

Purpose of the Study:

• To develop a multi-magnification convolutional neural network (CNN) for improved organ detection and segmentation in WSIs.
• To address limitations of single-magnification approaches in identifying complex organs.

Main Methods:

• A novel multi-magnification CNN, MMO-Net, was developed to integrate context and cellular details from various magnifications.
• The model was trained and evaluated on 320 WSIs from three contract research organization (CRO) laboratories.
• Performance was assessed for seven rat organs: liver, kidney, thyroid, parathyroid, urinary bladder, salivary gland, and mandibular lymph node.

Main Results:

• MMO-Net achieved state-of-the-art performance in organ detection and segmentation across multiple CROs.
• High Area Under the Receiver Operating Characteristic curve (AUROC) values (0.99-1.0) and Dice scores (≥0.9 for most organs) were reported.
• Excellent generalizability was demonstrated at both inter- and intra-CRO levels, with specific attention to separating adjacent organs like thyroid and parathyroid glands.

Conclusions:

• MMO-Net effectively localizes organs in histology images, offering a potential quality control for WSI metadata.
• The tool can serve as a preprocessing step for organ-specific artificial intelligence (AI) applications in toxicology.
• The study provides a publicly available dataset of WSIs and metadata to advance research in digital pathology.