Multi-task Semi-supervised Learning
for Vascular Network Segmentation
and Renal Cell Carcinoma Classification
Rudan Xiao
1(B)
, Damien Ambrosetti
2
, and Xavier Descombes
1
1
Universit´eCˆote d’Azur, Inria, CNRS, I3S, Nice, France
[email protected]
2
Hˆopital Pasteur, CHU Nice, Nice, France
Abstract.Vascular network analysis is crucial to define the tumoral
architecture and then diagnose the cancer subtype. However, automatic
vascular network segmentation from Hematoxylin and Eosin (H&E)
staining histopathological images is still a challenge due to the back-
ground complexity. Moreover, there is a lack of large manually anno-
tated vascular network databases. In this paper, we propose a method
that reduces reliance on labeled data through semi-supervised learning
(SSL). Additionally, considering the correlation between tumor classifica-
tion and vascular segmentation, we propose a multi-task learning (MTL)
model that can simultaneously segment the vascular network using SSL
and predict the tumor class in a supervised context. This multi-task
learning procedure offers an end-to-end machine learning solution to
joint vascular network segmentation and tumor classification. Experi-
ments were carried out on a database of histopathological images of
renal cell carcinoma (RCC) and then tested on both own RCC and open-
source TCGA datasets. The results show that the proposed MTL-SSL
model outperforms the conventional supervised-learning segmentation
approach.
Keywords:Vascular network segmentation
∙Semi-supervised
learning
∙Multi-task learning∙Renal cell carcinoma
1 Introduction
85% to 90% of kidney cancer are RCC, with the main subtypes being clear cell
RCC (ccRCC) with 75%, papillary RCC (pRCC) with 10% and Chromophobe
with 5% [11]. Currently, subtyping is essentially based upon pathological analy-
sis, consisting of cell morphology and tumor architecture [8]. [25] proved vascular
network analysis is important and relevant in RCC subtyping, however this clas-
sification work only used a few manually segmented vascular networks, which
limits its application potential. In this paper, we propose to build an automatic
vascular network segmentation model paired with a tumor classification scheme.
Data labeling is often the most challenging task. Labeling large-scale images
are laborious, time-consuming and exhibit low repeatability. This encouraged to
cffiThe Author(s), under exclusive license to Springer Nature Switzerland AG 2022
X. Xu et al. (Eds.): REMIA 2022, LNCS 13543, pp. 1–11, 2022.
https://doi.org/10.1007/978-3-031-16876-5
_1