Improved Myocardial Mapping by Cardiac MRI in Transthyretin Amyloidosis Cardiomyopathy Using Consistency-Parametrized AI Segmentation Model

Mateusz Cezary Florkow, Andreas Heindl, Majd Zreik, Ahmed Emam

Keywords: AI, segmentation, T1 mapping, ECV mapping, deep learning, guiding mask, cardiac MRI, MOLLI, medical image analysis

TL;DR: Improving the consistency of a segmentation across slices using a parametrized neural network

Abstract: Disease monitoring relies on accurate, standardised and consistent assessment of longitudinal data acquired over different time points. However, most current image segmentation approaches operate on individual images in isolation and do not leverage available information from other time points. In this work, we introduce a simple, parameterized loss function that incorporates a guiding mask to constrain image segmentation and promote consistency across repeated acquisitions in the same patient. The constraint can be adjusted according to image quality characteristics (e.g., signal-to-noise ratio, contrast). We trained a U-Net–based network using this loss and compared its performance against a baseline U-Net and SAM2. The approach was applied on a clinical dataset of transthyretin amyloid cardiomyopathy patients. Results indicate that the proposed model improved T1 mapping and extracellular volume estimation—a key marker for transthyretin amyloid cardiomyopathy.

Primary Subject Area: Segmentation

Secondary Subject Area: Application: Cardiology

Registration Requirement: Yes

Visa & Travel: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 84