ConferenceBest Paper Award, ICCT 2024

Refining Class Confusion with Pixel Adjacency and Continual Learning in Medical Image Segmentation

H. Lee, P. Puttimit, W. Thongsopa, S. Roh, Y. R. Lee, S. Y. Park, W. Y. Tak, S. K. Jung

Proceedings of the 7th International Conference on Culture Technology (ICCT 2024), pp. 86-92

Organs that touch in a CT scan get confused with each other. This work measures pixel adjacency between classes to pick the pairs a model mixes up, trains each pair as one merged class first, then separates them with continual segmentation using class-specific heads and CLIP text embeddings. Tested on three datasets: KNUH Body Composition, MM-WHS and FLARE22.

Download PDF

Results

DSC points on myocardium (MYO), MM-WHS: 80.19 to 90.46: +10.27DSC points on myocardium (MYO), MM-WHS: 80.19 to 90.46
average DSC on MM-WHS, up from a 84.81 baseline: 85.68%average DSC on MM-WHS, up from a 84.81 baseline
datasets, including 1,004 hospital CT slices: 3datasets, including 1,004 hospital CT slices

Per-class DSC on MM-WHS: baseline vs merging LVC with MYO, then separating

From Table III of the paper

U-Net baseline
Merge then separate (LVC + MYO)

DSC %

60708090100

LVC
+3.79
RVC
+1.34
LAC
−5.99
RAC
+0.47
MYO
+10.27
AA
+2.20
PA
−6.03
AVG
+0.87

Show as a table

Class	U-Net baseline	Merge then separate (LVC + MYO)	Δ
LVC	89.33	93.12	+3.79
RVC	84.10	85.44	+1.34
LAC	90.02	84.03	−5.99
RAC	81.63	82.10	+0.47
MYO	80.19	90.46	+10.27
AA	91.64	93.84	+2.20
PA	76.79	70.76	−6.03
AVG	84.81	85.68	+0.87

What we found

It works when the two classes have clearly different HU values: merging LVC with MYO lifted MYO by over 10 points and nudged neighbouring classes up too.
On KNUH Body Composition, IMAT (often mistaken for muscle) moved from 77.55 to 77.85; the average barely changed.
On FLARE22 the average dropped in every experiment (baseline 90.39 vs 87.96 to 89.46). With similar HU and a big size gap the model favours the larger organ, and organs whose shape changes a lot between slices are hard to separate.

Loading the paper…

Author's manuscript. Joint work with Kyungpook National University and Kyungpook National University Hospital.