Journal

Improving Cardiac Computed Tomography Scan Segmentation Using a U-Net Model with Continual Learning Techniques

W. Khamprapai, W. Thongsopa, C. Deejaiwong, J. Charoensuk, S. Mathulaprangsan, C. Chootong

Proceedings of Engineering and Technology Innovation, vol. 33, pp. 22-32, 2026

Takes the ICCT 2024 method to the heart. Seven cardiac structures are segmented from CT with U-Net Partial: adjacent organs are learned as merged classes first and separated with continual learning, with CLIP embeddings to tell classes apart, comparing optimizers and learning rates on MM-WHS 2017.

Download PDF DOI

Results

top DSC reported in the abstract: 94.08%top DSC reported in the abstract
top IoU reported in the abstract: 92.03%top IoU reported in the abstract
DSC for the PA + RVC group, best in the merging stage: 87.85%DSC for the PA + RVC group, best in the merging stage

Per-class DSC: U-Net baseline vs AdamW, lr 1e-4

From Table 3 of the paper; AdamW at 1e-4 was the best overall setting

U-Net baseline
U-Net Partial (AdamW 1e-4)

DSC %

60708090100

RAC
−0.22
LVC
−0.35
LAC
−7.29
MYO
+0.53
AA
+3.11
PA
−1.40
RVC
+0.86

Show as a table

Class	U-Net baseline	U-Net Partial (AdamW 1e-4)	Δ
RAC	81.63	81.41	−0.22
LVC	89.33	88.98	−0.35
LAC	90.02	82.73	−7.29
MYO	80.19	80.72	+0.53
AA	91.64	94.75	+3.11
PA	76.79	75.39	−1.40
RVC	84.10	84.96	+0.86

What we found

AdamW at a 1e-4 learning rate was the strongest setting overall, most clearly on the aorta (AA) at 94.75 DSC.
PA and LAC remain hard: PA is small and varies a lot between patients, and LAC has weak boundaries in CT. The method reduces confusion between large neighbours but can blur the edges of small structures.
Limits: one public dataset with few CT volumes, compared only against a U-Net baseline.

Loading the paper…