MiThyCA
Microscopic foci of papillary Thyroid Carcinoma-like nuclear features identification with AI in Whole-Slide Images.
MiThyCA is an advanced computational pathology pipeline designed to assist pathologists in the rapid and automated detection of subcentimeter Papillary Thyroid Carcinoma (PTC) foci within Whole-Slide Images (WSIs).
By utilizing a tandem deep-learning architecture, MiThyCA identifies neoplastic areas and detects "sprinkling" areas—tiny, abrupt nuclear alterations that are often time-consuming to find manually. This tool significantly reduces turnaround times, processing a full slide in as little as 11 seconds.
Tandem AI Architecture
MiThyCA employs a sequential “two-step” approach to maximize both sensitivity and computational efficiency:
- Model 1 (Neoplasm Finder): A Convolutional Neural Network (CNN) that screens the entire WSI to identify neoplastic or abnormal tissue regions.
- Model 2 (PTC-like Feature Detector): A Vision Transformer (TinyViT) that performs a high-resolution analysis only on the areas identified by Model 1, specifically searching for PTC-like nuclear features (e.g., nuclear clearing, grooves, and pseudoinclusions).
Key Results & Performance
The pipeline was validated on a multi-institutional cohort, demonstrating high reliability across various thyroid lesions including NIFTP, PTC, and metastatic lymph nodes.
- Clinical Value: Effectively detects the “sprinkling sign” in NIFTP and identifies metastatic micro-foci in lymph nodes that might be overlooked during routine manual screening.
Technical Highlights
Tandem Strategy
Combines the efficiency of CNNs with the high-context attention of Vision Transformers (ViT).
Hardware Agnostic
Optimized to run on standard hospital workstations even without high-end dedicated GPUs.
Explainable Heatmaps
Integrated with QuPath to provide visual heatmaps, keeping the pathologist "in-the-loop."