Amit Mehndiratta1, Esha Badiya Kayal1, Raju Sharma2, Sameer Bakhshi3, and Devasenathipathy Kandasamy2
1Centre for Biomedical Engineering, Indian Institute of Technology Delhi, New Delhi, India, 2Department of RadioDiagnosis, All India Institute of Medical Sciences, New Delhi, India, 3Dr. BRA Institute-Rotary Cancer Hospital, All India Institute of Medical Sciences, New Delhi, India
Synopsis
Accuracy and consistency in RECIST(Response evaluation
criteria in solid tumors) measurements are crucial as it directly impacts
patient treatment options. Manual RECIST
measurement, requiring high expertise & attention, is
time-expensive, prone-to-error, operator-subjective. we
propose an automated tumor segmentation and RECIST score estimation method that
uses MRI image slices as input, delineates the tumor in 3D, identifies the MRI
slice with maximum-tumor-burden and then measures the tumor-diameter and RECIST1.1
score for treatment response assessment. Proposed method produced reliable
and reproducible automated RECIST score measurements in current bone tumor dataset and might be
useful as decision support tool saving manual-effort and
reading-time.
Purpose
Response evaluation criteria in solid tumors
(RECIST) is the standard method to measure and score the tumor growth rates
across different time points to evaluate treatment responses in cancer patients
(1). Accuracy
and consistency in RECIST measurements are crucial as it directly impacts
patient treatment options. Manual RECIST
measurement, requiring high expertise & attention, is
time-expensive, prone-to-error and operator-subjective. The purpose of this study was to develop a computer
aided diagnostic (CAD) tool for automated tumor segmentation and RECIST score
estimation with reasonable
accuracy, consistency and speed.Methods
Dataset: To
implement and test the performance of proposed CAD tool, bone tumor data of forty
patients (N=40;Male:Female=30:10;Age=17.7±5.9years) with biopsy
proven Osteosarcoma were used. For each patient diffusion weighted MRI (DWI) was
acquired using 1.5T Phillips Achieva MRI scanner before treatment (baseline) and after
completion of neoadjuvant chemotherapy (NACT) (follow-up, after 10-12
weeks). DWI was acquired using free breathing Spin-Echo-Echo-Planar
Imaging (SP-EPI) with TR/TE=7541/67msec, matrix-size=192×192,
field-of-view=250´250mm2, slice-thickness/Gap=5.0mm/0.5mm,
voxel-size=1.3/1.3/5.0mm, b-value=0-800s/mm2
with 64 axial-slices.
RECIST1.1 score: To assess RECIST1.1 score, tumor-diameter
is manually measured in the maximum cross-sectional area of the tumor at
different time-points in the course of treatment and the changes in tumor
burden is evaluated and scored according to RECIST1.1 criteria. RECIST1.1
scores are defined as, Complete-response (CR): total disappearance of tumor;
Partial-response (PR): Minimum 30% decrease in tumor-diameter;
Progressive-disease (PD): minimum 20% or 5 mm absolute increase in
tumor-diameter; Stable-disease (SD): neither PR nor PD.
Ground-Truth: Tumor-diameter (in cm) was
measured manually on axial b=800s/mm2 DWI slice (DWI800)
having the maximum cross-sectional area of tumor. Tumor-diameter was measured
using RadiAnt DICOM Viewer 1.9.16version software (www.radiantviewer.com) and the
slice-number with maximum tumor burden (max-burden-sliceno) was noted.
Manual segmentation and demarcation of tumor Region-of-interest (ROI) was
performed on each DWI800 image covering the whole tumor
using MRIcron 8/2014 version software (www.nitrc.org). Tumor-volume (in cc) was
calculated separately at baseline and follow-up for each patient by
accumulating the tumor ROIs across all slices. Illustrative examples of original T2W fat-saturated
image and DWI800 image from a representative patient with Osteosarcoma with delineated
tumor-diameter and tumor ROI are depicted in Fig1.a, Fig1.b and Fig1.c
respectively. Relative percentage changes in tumor-diameter and
tumor-volume from baseline to follow-up were calculated and RECIST1.1 score was
evaluated for each patient.
Proposed CAD Tool: 3D image stack of DWI800 were pre-processed by morphological
operations to focus only on target anatomical part and were normalized
to total 256 grey levels (0-255) to facilitate further operation. Automatic tumor segmentation in
3D was performed using unsupervised clustering based algorithm Simple-linear-iterative-clustering
Superpixels (SLIC-S)(2) (Fig.2). Superpixels were generated by clustering
voxels based on intensity similarity and proximity in plane. Experimentally the
number-of-supervoxels was set to 20 with the compactness of 0.025 and 50
number-of-iterations. Mean intensity of supervoxels were calculated and
Histogram analysis was performed in whole tumor volume and four Ostu-thresholds
were estimated considering multi-Gaussian distribution. Supervoxels above
threshold-3 were selected experimentally and merged followed by morphological
operations which gave the best segmentation results. Using connected
component analysis (3) length of the
major-axis (longest cross-sectional axis) of all components of tumor in each
slice were determined and added automatically to obtain the tumor-diameter for
that slice. Maximum of the tumor-diameters from all slices was selected programmatically
and multiplied with voxel-length 0.13cm to obtain the final tumor-diameter (in
cm) and the associated slice-number was noted as max-burden-slice_no. Tumor-volume
(in cc) was determined automatically by multiplying the number of voxels in
segmented whole tumor region with the voxel-volume 1.3x1.3x5.0x10-3cc. Relative-percentage-changes
in tumor-sizes across time-points were scored using RECIST1.1 criteria.
Accuracy metric: Segmentation accuracy
was estimated by Dice-coefficient(DC),
Jaccard-Index(JI), Precision(P) and Recall(R). Evaluated Apparent-diffusion-coefficient(ADC), tumor-diameter, max-burden-sliceno and
tumor-volume in segmented tumor-mask and ground-truth
tumor-mask were compared using paired-t-test
for statistical significance(p<0.05)
and Pearson-correlation-coefficient(PCC).
Misclassification error rate (MER) for automated scoring methods for
classifying patients in different response group was evaluated as $$MER= \frac{(Total no.of misclassification)}{(Total no.of patients)}$$.
Proposed tool was implemented using
libraries in Matlab2015b, Philadelphia, USA. Statistical analysis for accuracy
calculation was performed using SPSS 16.0 software.Results
Illustrative
examples of tumor segmentation are represented in Fig2.A and Fig2.B from a representative patient. Qualitative
SLIC-S produced satisfactory segmentation of tumor in comparison to the ground
truth tumor ROI at baseline and follow-up. Table1 shows segmentation accuracy for
all patients and average ADC in
tumor-volume. At baseline high accuracy was observed (DC:~83%;JI:~72%;P:~83%;R:~86%); while at follow-up it was satisfactory (DC:~72%;JI:~64%;P:~63%;R:~86%). Mean ADC (1.29-1.31x103mm2/s) values in segmented
masks were not significantly different (p>0.05)
and showed excellent correlation (PCC=0.85-0.89) with the ADC(1.3±0.33x103mm2/s) values of ground-truth
mask. Table2 depicted the RECIST1.1 score and volumetric response score in
patient cohort using manual and automated methods. Miss-classification rate for
proposed CAD tool was for RECIST1.1 and volumetric response score were 15%
and 26% respectively. Assessment times
were 2-3sec using Intel Xeon CPU E3-1241v3@3.50
GHz processor and 32 GB RAM. Conclusion
The automation of RECIST response evaluation
provided considerable saving in manual effort and reading time with reliable
and reproducible accuracy. Proposed CAD tool produced promising segmentation and RECIST
score measurements in current bone tumor dataset and might be useful as decision
support tool helping physicians in improving oncologic readings.Acknowledgements
Authors would like to thanks the nurse and support staff of AIIMS for support in data acquisition and scanning. References
1. Eisenhauer EA, Therasse P, Bogaerts J, et al.: New
response evaluation criteria in solid tumours: revised RECIST guideline
(version 1.1). Eur J Cancer 2009; 45:228–247.
2. Achanta R, Shaji A, Smith K,
Lucchi A, Fua P, Susstrunk S: SLIC Superpixels. IEEE Trans Pattern Anal Mach
Intell 2012; 34:2274–2282.
3. He L, Ren X, Gao Q, Zhao X, Yao B,
Chao Y: The connected-component labeling problem : A review of state-of-the-art
algorithms. Pattern Recognit 2017; 70:25–43.