Recurrent Mask Refinement for Few-Shot Medical Image Segmentation

Hao Tang Xingwei Liu Shanlin Sun Xiangyi Yan Xiaohui Xie
Department of Computer Science, University of California Irvine
University of California Irivne, Irvine, California, 92697
{htang6, xingweil, shanlins, xiangyy4, xhx}@uci.edu

1 Supplementary #1

Regarding the selection of support and query images, we used randomly chosen pairs of support and query images, instead of one fixed support, in order to make sure that our results are not biased by a particular choice of support/query pairs in Table 1. Among the three datasets (ABD-110, ABD-30 and ABD-MR) we tested, the selection of support/query images on ABD-110 are exactly the same for all compared methods. We ensured the same sequence of support/query pairs (albeit random) were used by fixing the seed of the random number generator. Our selection of support/query images was different for ABD-30 and ABD-MR from [1], which used a fixed support image. To ensure the results are directly comparable, we reran our method following exactly the same selection of support as in [1]. Our conclusion remains the same (Supplementary Table 1). Our result (labeled “RP-Net (fixed support)”) shows the proposed method still outperforms the SOTA method by a large margin.

Dataset	Method	Spleen	Kidney L	Kidney R	Liver	mean
ABD-110	PANet-init [3]	30.95 $\pm$ 1.09	19.24 $\pm$ 0.37	17.64 $\pm$ 0.71	49.91 $\pm$ 0.34	29.43
	PANet [3]	35.89 $\pm$ 1.75	40.22 $\pm$ 1.71	41.54 $\pm$ 0.82	52.36 $\pm$ 0.60	42.50
	SE-Net [2]	29.48 $\pm$ 1.07	37.48 $\pm$ 2.08	37.53 $\pm$ 1.97	19.09 $\pm$ 0.36	30.89
	SSL-ALPNet [1]	64.90 $\pm$ 1.62	61.58 $\pm$ 2.53	64.05 $\pm$ 2.27	71.83 $\pm$ 1.81	65.59
	Affine	50.42 $\pm$ 0.91	53.04 $\pm$ 1.57	52.025 $\pm$ 2.17	66.99 $\pm$ 1.20	55.62
	RP-Net (Ours)	78.77 $\pm$ 0.64	81.89 $\pm$ 1.45	85.12 $\pm$ 0.98	81.88 $\pm$ 0.63	81.91
ABD-30	SE-Net [2]	0.23	32.83	14.34	0.27	11.91
	PANet-init [3]	23.82	13.97	14.17	50.27	25.55
	PANet [3]	25.59	32.34	17.37	38.42	29.42
	SSL-ALPNet [1]	60.25	63.34	54.82	73.65	63.02
	Affine	48.99 $\pm$ 1.48	43.44 $\pm$ 2.04	45.67 $\pm$ 1.45	68.93 $\pm$ 0.88	51.75
	RP-Net (Ours)	69.85 $\pm$ 2.34	70.48 $\pm$ 2.55	70.00 $\pm$ 0.89	79.62 $\pm$ 0.91	72.48
	RP-Net (fixed support)	68.27	71.59	70.27	80.51	72.66
ABD-MR	SE-Net [2]	51.80	62.11	61.32	27.43	50.66
	PANet-init [3]	34.59	18.63	22.50	47.43	30.78
	PANet [3]	50.90	53.45	38.64	42.26	46.33
	SSL-ALPNet [1]	67.02	73.63	78.39	73.05	73.02
	Affine	62.87 $\pm$ 1.80	64.70 $\pm$ 4.71	69.10 $\pm$ 1.15	65 $\pm$ 1.65	65.41
	RP-Net (Ours)	76.35 $\pm$ 0.66	81.40 $\pm$ 2.10	85.78 $\pm$ 1.12	73.51 $\pm$ 1.55	79.26
	RP-Net (fixed support)	75.96	80.18	86.63	73.52	78.99

Table 1: DSC comparison with other methods on ABD-110, ABD-30 and ABD-MR (unit: %).

References

[1] Cheng Ouyang, Carlo Biffi, Chen Chen, Turkay Kart, Huaqi Qiu, and Daniel Rueckert. Self-supervision with superpixels: Training few-shot medical image segmentation without annotation. In European Conference on Computer Vision, pages 762–780. Springer, 2020.
[2] Abhijit Guha Roy, Shayan Siddiqui, Sebastian Pölsterl, Nassir Navab, and Christian Wachinger. ‘squeeze & excite’guided few-shot segmentation of volumetric images. Medical image analysis, 59:101587, 2020.
[3] Kaixin Wang, Jun Hao Liew, Yingtian Zou, Daquan Zhou, and Jiashi Feng. Panet: Few-shot image semantic segmentation with prototype alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9197–9206, 2019.