AI-assisted Diagnosis of Endometrium (Pre)malignancies in Pipelle Biopsies
Keywords
Loading...
Authors
Issue Date
2022-11-29
Language
en
Document type
Journal Title
Journal ISSN
Volume Title
Publisher
Title
ISSN
Volume
Issue
Startpage
Endpage
DOI
Abstract
Endometrial carcinoma (EC) is the third most common cause of death in cancers that only affect women. The
most frequent treatment for endometrial carcinoma is the removal of the uterus (abdominal hysterectomy).
This treatment is less radical than it seems, since the cancer often occurs in postmenopausal women,
where the function of the uterus has become obsolete. However, the desire to preserve the uterus, both in
postmenopausal and non-postmenopausal women, will require different treatment options. The outcome
of treatment options such as radiation therapy, hormone therapy or chemotherapy will become more
favorable with early diagnosis of the disease. The need for early diagnosis caused a series of innovations of
endometrial sampling devices to decrease the invasiveness and increase the specificity and sensitivity. The
most used device today is the Pipelle de Cornier, which has high sensitivity and specificity in diagnosing
EC. This elongated suction device has been researched clinically, but to date no studies have investigated
the potential for computed-aided automation. The way the Pipelle uses vacuumto extract tissue does not
yield usable material in 20-30% of the cases. Combined with the vast amount of benign samples, since the
non-invasiveness will have specialists opt for it faster, results in a high number of cases that are less relevant
for a pathologist.
The aimof this study was, therefore, to investigate the feasibility of computer-aided screening systems
on Pipelle sampled tissue and developing, to the best of our knowledge, the first deep learning system to
address this problem. The overall aim of this project was to develop a computer systemthat can improve
the diagnostic procedure of Pipelle sampled endometrial tissue to reduce the amount of benign and noninformative
samples a pathologist would have to diagnose, such that the pathologist can focus on the difficult
cases.
Two common computer vision techniques were explored: weakly-supervised slide-level classification
and pixel-level semantic segmentation. Multiple experiments have been conducted on both techniques to
ensure the best fit on the data and to test our novel contributions to each. The self-configuring nnU-Net has
been tested on and tailored to pathology imagery by proposing a multi-branch multi-resolution architecture,
showcasing qualitative improvements. The weakly-supervised multiple-instance learning architecture CLAM
(Clustering-constrained AttentionMultiple Instance Learning) has been extended to allowprior knowledge, in
the formof clinical information, to be incorporate into the network and showed quantitative improvements
over the baseline. An ensemble of the CLAM architecture, hosted on grand-challenge.org1, was able to
achieve a sensitivity of 100% at a false positive rate of 10%, potentially reducing the amount of work by 72.9%.
Description
Citation
Supervisor
Faculty
Faculteit der Sociale Wetenschappen