Some problems are easier to spot in motion than at a standstill. After a heart attack, persisting damage to tiny blood vessels can block blood flow to the cardiac muscles. Called microvascular obstruction (MVO), this condition can be difficult to distinguish in a snapshot.
Cardiac magnetic resonance (CMR) is the standard for visualising heart tissue structures, with a chemical called gadolinium often used to enhance image contrast. But gadolinium-based dyes can pose risks for patients with impaired kidney functions, creating a need for contrast-free alternatives like Cine CMR.
“Although MVO may not look dramatically different from the surrounding tissue based on a single Cine image, the affected cardiac muscles exhibit altered motion patterns that Cine CMR can capture over a series of frames,” said Xulei Yang, a Principal Scientist at the A*STAR Institute of Advanced Intelligence and Computing (A*STAR IAIC).
This shaped the approach taken by Yang and A*STAR IAIC colleagues who worked with researchers at the Mechanobiology Institute, National University of Singapore; and the National Heart Centre Singapore. The researchers hypothesised that they could leverage artificial intelligence (AI) to extract new clues from cardiac motion patterns in Cine CMR sequences, hoping to overcome the limitations of relying on structural information from individual Cine CMR frames for MVO detection even without contrast enhancement.
Accordingly, the team built a spatiotemporal-sensitive network with a dual-encoder design. Acting like two sets of eyes, the encoders separately captured the heart’s structural characteristics and its movement features, the latter processed as image intensity differences between consecutive Cine CMR frames. The complementary signals were then combined using a guided decoder.
“When the model is deciding whether to flag a particular tissue area, the motion information helps guide its attention towards structural regions whose behaviour looks unusual,” said Yang Yu, a former Senior Researcher at A*STAR IAIC.
After training and testing using 621 pairs of Cine and gadolinium-enhanced CMR scans from 125 cases, the model achieved a Dice score of 0.5556, reflecting how closely its MVO detection matched expert analysis, compared to 0.3271 using only the structural encoder alone or 0.4636 with motion information alone.
“Medical imaging contains much more information than simply brightness, colour or clearly visible boundaries. In Cine CMR, information is also encoded in how anatomy changes over time. Our work supports the importance of combining these two sources of information,” Yu noted.
While the approach showed promising results, large-scale clinical validation studies, involving more diverse patient groups with varying MVO size and severity levels, are needed before it can become a standalone diagnostic tool. Moreover, the researchers may have to adapt the model to analyse datasets from hospitals using different CMR scanner technologies and imaging protocols. “I would describe our work as an important step towards contrast-free MVO assessment, rather than a replacement for standard contrast-enhanced CMR imaging today,” Yang remarked.
The A*STAR-affiliated researchers contributing to this research are from the A*STAR Institute of Advanced Intelligence and Computing (A*STAR IAIC).
