Progressively rened deep joint registration segmentation ProRSeg of gastrointestinal organs at risk Application to MRI and cone-beam CT

2025-05-02 0 0 2.09MB 24 页 10玖币

侵权投诉

Progressively reﬁned deep joint registration segmentation

(ProRSeg) of gastrointestinal organs at risk: Application to

MRI and cone-beam CT

Jue Jiang1,†, Jun Hong1,?, Kathryn Tringale2, Marsha Reyngold2, Christopher Crane2, Neelam Tyagi1,

Harini Veeraraghavan1,†

Department of Medical Physics1, Memorial Sloan Kettering Cancer Center, 1275 York Av-

enue, New York, NY 1006

Department of Radiation Oncology2, Memorial Sloan Kettering Cancer Center, 1275 York

Avenue, New York, NY 1006

†: These authors contributed equally

?: work performed when author was at MSKCC

Corresponding Author Address: Box 84 - Medical Physics, Memorial Sloan Kettering Cancer

Center, 1275 York Avenue, New York, NY 10065 .

Corresponding Author Email: veerarah@mskcc.org

Abstract

Background: Adaptive radiation treatment (ART) for locally advanced pancre-

atic cancer (LAPC) requires consistently accurate segmentation of the extremely mo-

bile gastrointestinal (GI) organs at risk (OAR) including the stomach, duodenum, large

and small bowel. Also, due to lack of suﬃciently accurate and fast deformable image

registration (DIR), accumulated dose to the GI OARs is currently only estimated, fur-

ther limiting the ability to more precisely adapt treatments.

Purpose: Develop a 3-D progressively reﬁned joint registration-segmentation

(ProRSeg) deep network to segment and align treatment MRIs, then evaluate segmen-

tation accuracy, registration consistency, and feasibility for OAR dose accumulation.

Method: ProRSeg was trained using 5-fold cross-validation with 110 T2-weighted

MRI acquired at 5 treatment fractions from 10 diﬀerent patients, taking care that

same patient scans were not placed in training and testing folds. Segmentation ac-

curacy was measured using Dice similarity coeﬃcient (DSC) and Hausdorﬀ distance

at 95th percentile (HD95). Registration consistency was measured using coeﬃcient of

variation (CV) in displacement of OARs. Ablation tests and accuracy comparisons

against multiple methods were done. Finally, applicability of ProRSeg to segment

cone-beam CT (CBCT) scans was evaluated on 80 scans using 5-fold cross-validation.

Results: ProRSeg processed 3D volumes (128 ×192 ×128) in 3 secs on a NVIDIA

Tesla V100 GPU. It’s segmentations were signiﬁcantly more accurate (p < 0.001)

than compared methods, achieving a DSC of 0.94 ±0.02 for liver, 0.88±0.04 for large

bowel, 0.78±0.03 for small bowel and 0.82±0.04 for stomach-duodenum from MRI.

ProRSeg achieved a DSC of 0.72±0.01 for small bowel and 0.76±0.03 for stomach-

duodenum from CBCT. ProRSeg registrations resulted in the lowest CV in displace-

ment (stomach-duodenum CVx: 0.75%, CVy: 0.73%, and CVz: 0.81%; small bowel

arXiv:2210.14297v1 [eess.IV] 25 Oct 2022

CVx: 0.80%, CVy: 0.80%, and CVz: 0.68%; large bowel CVx: 0.71%, CVy: 0.81%,

and CVz: 0.75%). ProRSeg based dose accumulation accounting for intra-fraction

(pre-treatment to post-treatment MRI scan) and inter-fraction motion showed that

the organ dose constraints were violated in 4 patients for stomach-duodenum and for

3 patients for small bowel. Study limitations include lack of independent testing and

ground truth phantom datasets to measure dose accumulation accuracy.

Conclusions: ProRSeg produced more accurate and consistent GI OARs segmenta-

tion and DIR of MRI and CBCTs compared to multiple methods. Preliminary results

indicates feasibility for OAR dose accumulation using ProRSeg.

Keywords: Recurrent deep networks, GI organs, segmentation, registration, MRI,

CBCT.

I. Introduction

MR-guided adaptive radiation therapy (MRgART) is a new treatment that allows for ra-

diative dose escalation of locally advanced pancreatic cancers (LAPC) with higher precision

than conventionally used cone-beam CTs (CBCT) due to improved soft-tissue visualiza-

tion on MRI. MR-LINAC treatments also allow daily treatment adaptation and replanning

to account for the changing anatomy. Anatomy changes result from day-to-day variations

in organ shape and conﬁguration as well as motion due to peristalsis and breathing, all

of which introduce large geometric uncertainties to the delivery of radiation. However,

widespread adoption of MRgART is hampered by the need for manual contouring and plan

re-optimization, which together can take 40 to 70 mins8,18 daily. Hence, there is a clinical

need for consistently accurate and fast auto-segmentation of the gastrointestinal (GI) organs

at risk (OARs) including the stomach, duodenum, small and large bowel.

Highly accurate segmentation, measured as a Dice similarity coeﬃcient (DSC) exceeding 0.8

of abdominal organs such as the liver, kidneys, spleen, as well as the stomach (excluding

duodenum) has been reported by using oﬀ-the-shelf deep learning (DL) architectures includ-

ing nnUnet19 and Unet26, as well as customized dense V-Nets14 and new transformer based

methods33 applied to CT images. Slice-wise priors provided as manual segmentations32,

multi-view methods using inter-slice information from several slices and dense connections33

as well as self-supervised learning of transformers21 have shown the ability to segment the

more challenging GI OARs such as large and small bowel and duodenum from MRI. How-

ever, the need for manual editing32 and large number of adjacent slices33 required to provide

priors may reduce the number of available training sets and the practicality (due to need for

manual editing) of such methods.

Besides segmentation, reliable deformable image registration (DIR) is also needed for voxel-

wise OAR dose accumulation in order to ensure that the prescription dose was delivered

to the targets while sparing organs of unnecessary radiation. DIR based contour propaga-

tion23,30,34 is one approach to simultaneously solve both deformable dose accumulation and

segmentation. This is also a convenient option as DIR methods are commonly available

techniques in commercial software platforms. However, commonly available DIR methods

often use small deformation frameworks based on parameterizing a displacement ﬁeld added

to an identity transform, which cannot preserve topology4for large organ displacements.

Deep learning image registration (DLIR) methods6,12,17,31 are often faster and more accurate

than iterative registration methods because they directly compute the diﬀeomorphic trans-

formation between images in a single step instead of solving a non-linear optimization to

align every image pair. DLIR methods, which typically use stationary velocity ﬁeld (SVF)

to compute the diﬀeomorphic transformation, reduce the computational requirements by

reducing the search space to a set of diﬀeomorphisms that are within a Lie structure, but

it also limits their ﬂexibility in handling large and complex deformations24. Furthermore,

DLIR network optimization as well as iterative registration optimization typically focuses

on minimizing an energy function composed of global smoothness and a variational intensity

regularization, which is insuﬃcient to handle abrupt and large diﬀerences in motion occur-

ring between organs, especially at the organ boundaries13.

Compositional DIR strategies that extract diﬀeomorphic transformation in stages such as

used for non-sliding and sliding organs13, adaptive anisotropic ﬁltering of the incrementally

reﬁned deformation vector ﬁeld (DVF)25 as well as cascaded network formulations4,10,35

are more robust than the single step methods for handling large deformations while still

retaining the SVF assumption, thus providing computational speed up compared to the

time-dependent diﬀeomorphic registration methods. However, cascaded DLIR methods are

limited by the memory requirements and thus require sequential training of individual net-

works, which increases training time. Additionally, because the networks in the cascade

are trained one after another, there is no guarantee that the deformations modeled in the

prior steps will be retained in the future steps. Recurrent registration method (R2N2)

that computes local parameterized Gaussian deformations27 has demonstrated ability to

model large anatomic deformations occurring in a respiratory cycle. However, the use of

local parametrization restricts the ﬂexibility of this approach to handle large and contin-

uous deformations. R2N2 was shown to be less accurate than a progressive registration

method computing a continuous deformation ﬂow ﬁeld for quantifying longitudinal tumor

volume changes22. Our approach improves on these works to compute topology preserving

(quantiﬁed by non-negative Jacobian determinant) diﬀeomorphic deformations and multi-

organ segmentations using a progressive joint registration-segmentation (ProRSeg) approach,

wherein deformation ﬂow computed at a given step is conditioned on the prior step using a

3D convolutional long short term memory network (CLSTM)28. Because the DVF is mod-

eled as a continuous and diﬀerentiable ﬂow-ﬁeld, it is invertible, thus ensuring diﬀeomorphic

transformations. ProRSeg is optimized using a multi-tasked learning of a registration and

segmentation network, which allows it to leverage the implicit backpropagated errors from

the two networks. Multi-tasked networks have previously shown to produce more accurate

normal tissue segmentation than individually trained DL networks7,12,17,31.

ProRSeg is most similar to a prior registration-segmentation method that we developed for

tracking lung tumors22 from cone-beam CT (CBCT) images. However, ProRSeg accounts for

both respiratory and large organ shape variations, while our prior work was only concerned

with tracking linearly shrinking tumors during radiation treatment. ProRSeg aligns images

with large deformation by computing a smooth interpolated sequence of dense deformation

ﬂows using 3D CLSTM28 networks implemented in the encoder layers of registration and

segmentation networks. The CLSTM explicitly enforces consistency between the individ-

CLSTM

Warp

CLSTM

......

Warp

......

DVF DVF

Fixed Image

Moving Image

Hidden State

CLSTM

......

CLSTM

Conv

DVF

CLSTM

Conv

CNN

Conv

CLSTM

Conv

Seg

(a) (b)

(c)

RRN

RSN

Figure 1: (a) Schematic of recurrent Registration Network (RRN), where convolutional (Conv)

layers in the encoder are combined with 3D-CLSTM. (b) Recurrent Segmentation Network (RSN)

uses a Unet-3D backbone with 3D-CLSTM placed after convolutional blocks in the encoder layers.

layers for progressively reﬁning the registration and segmentation are shown. RSN combines xt

with the progressively aligned images xi=1,...N

mand segmentations yi=1,...N

mproduced by RRN as

inputs to its CLSTMs to generate segmentation ytin N steps.

ual steps and conditions the deformations computed in subsequent steps on the prior steps,

which allows incremental reﬁnement of spatial alignment without destroying prior alignment.

CLSTM formulation employs convolutional ﬁlters, thereby allowing for spatially continuous

and diﬀerentiable DVF formulation. Second, the segmentation network uses progressively

aligned spatial appearance and geometry priors pertaining to previous treatment fraction

(produced by the registration network) as inputs to constrain the segmentations. Hence, the

segmentation network avails information about the organs and their geometry from a prior

treatment fraction, which increases robustness to arbitrary variations occurring in the GI

organs. In contrast, prior works12,31 used a weaker regularizing constraint to ensure that the

outputs of registration and segmentation networks matched as additional losses during train-

ing. Third, segmentation consistency loss enforcing similarity of segmentation produced by

registration network, the segmentation network and the expert provides supervised feedback

to both networks. We show such a loss improves accuracy.

Our contributions are: (i) a simultaneous registration-segmentation approach for segment-

ing GI OARs from MRI while computing voxel-wise deformable dose accumulation, (ii) use

of registration derived spatially aligned appearance and geometry priors to constrain seg-

mentation that increases accuracy, (iii) use of a 3D CLSTM implemented in the encoders

of both registration and segmentation networks that increases robustness to arbitrary or-

gan deformations by modeling such deformations as a progressively varying dense ﬂow ﬁeld.

(iv) Finally, we evaluated ProRSeg for segmenting GI organs (stomach-duodenum and small

bowel) from treatment CBCTs.

II. Materials and Method

II.A. Pancreas MRI dataset for MR-MR registration-segmentation

The retrospective analysis was approved by the institutional internal review board. One

hundred and ten 3D T2-weighted MRIs acquired from on treatment MRIs from 10 patients

undergoing ﬁve fraction MR-guided SBRT to a total dose of 50 Gy were analyzed. A pneu-

matic compression belt set according to the patient convenience was used to minimize gross

tumor volume (GTV) and GI organs motion occurring within 5mm of the GTV as described

in our prior study29. The dose constraints to GI organs were deﬁned as Dmax or D0.035cm3≤

33 Gy and D5cm3≤25 Gy. D5cm3for large bowel was ≤30 Gy. In each treatment fraction,

three 3D T2-weighted MRI (TR/TE of 1300/87 ms, voxel size of 1 ×1×2 mm3, FOV of

400 ×450 ×250 mm3) were acquired at pre-treatment, veriﬁcation, and following treatment

on that fraction called post-treatment MRI. Veriﬁcation MRI was acquired immediately be-

fore beam on to conﬁrm patient anatomy had not shifted during contouring and adaptive

replanning. Six patients had pre-treatment, veriﬁcation, and post-treatment MRI with seg-

mentation on all ﬁve fractions with the remaining 4 containing only pre-treatment MRI for

all fractions. Additional details of treatment planning are in prior study1,29.

Expert contouring details: Stomach-duodenum, large bowel, and small bowel, as well

as liver were contoured on all the available treatment fraction MRIs by an expert medical

student and veriﬁed by radiation oncologists, and represented the ground truth for verifying

the ProSeg segmentations and deformable image registration (DIR).

II.B. Pancreas dataset for pCT-CBCT registration-segmentation

ProRSeg was additionally evaluated on 80 CBCT scans acquired from 40 patients with

LAPC and treated with hypofractionated RT on a regular linac (15 to 20 fractions with

daily CBCTs used for guidance) from an institutional IRB approved retrospective research

protocol and used in a prior work15. For the purpose of study, a planning CT (pCT) and

two CBCT scans acquired on diﬀerent days in a deep inspiration breath hold (DIBH) state

using an external respiratory monitor (Real-time Position Monitor, Varian Medical Systems)

were used. pCT scans were acquired in DIBH with a diagnostic quality scanner (Brilliance

Big Bore, Philips Health Systems; or DiscoveryST, GE Healthcare). The kilovoltage CBCT

scans were acquired with 200-degree gantry rotation. The CBCT reconstruction diameter

was 25 cm and length was 17.8 cm.

Expert contouring details: Both pCT and CBCT scans were delineated by a radiation

oncologist to provide stomach and ﬁrst two segments of the duodenum and the remainder

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

Progressivelyreneddeepjointregistrationsegmentation(ProRSeg)ofgastrointestinalorgansatrisk:ApplicationtoMRIandcone-beamCTJueJiang1;y,JunHong1;?,KathrynTringale2,MarshaReyngold2,ChristopherCrane2,NeelamTyagi1,HariniVeeraraghavan1;yDepartmentofMedicalPhysics1,MemorialSloanKetteringCancerCenter,1275Yo...

展开>> 收起<<

Progressively rened deep joint registration segmentation ProRSeg of gastrointestinal organs at risk Application to MRI and cone-beam CT.pdf

共24页,预览5页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Progressively rened deep joint registration segmentation ProRSeg of gastrointestinal organs at risk Application to MRI and cone-beam CT

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: