ISO/IEC 3532-2:2024

Information technology - Medical image-based modelling for 3D printing - Part 2: Segmentation

This document provides an overview of the segmentation process for medical image-based modelling of human bone. This document specifies a standardized process to improve the performance of human bone segmentation. This document is also applicable to medical 3D printing systems that include medical 3D modelling capabilities.

Technologies de l'information — Modélisation médicale à base d'images pour l'impression 3D — Partie 2: Segmentation

General Information

• Status: Published
• Publication Date: 12-Feb-2024
• ICS: 25.030 - Additive manufacturing
• ICS: 35.240.80 - IT applications in health care technology
• Technical Committee: ISO/IEC JTC 1 - Information technology
• Drafting Committee: ISO/IEC JTC 1/WG 12 - 3D Printing and scanning
• Current Stage: 6060 - International Standard published
• Start Date: 13-Feb-2024
• Due Date: 24-May-2024
• Completion Date: 13-Feb-2024

Overview

ISO/IEC 3532-2:2024 - Information technology - Medical image‑based modelling for 3D printing - Part 2: Segmentation - defines a standardized process for medical image segmentation of human bone used in medical 3D modelling and 3D printing systems. The standard covers the end‑to‑end segmentation workflow (data preparation, preprocessing, annotation, model selection, evaluation, deployment and post‑processing) and aims to improve performance, consistency and traceability of bone segmentation from CT and MR images.

Key topics and requirements

• Scope: Focus on segmentation for human bone and applicability to medical 3D printing systems that include modelling capabilities.
• Segmentation process: Seven main steps are specified - data preparation, preprocessing for segmentation, annotation, selection of segmentation network model, performance evaluation, model deployment/running, and post‑processing.
• Imaging considerations: Guidance on medical image acquisition and reconstruction (CT, MR), with annexes addressing CT scanning for orbital bone and characteristics of orbital bone segmentation.
• Preprocessing: Intensity normalization and spacing normalization to standardize inputs across datasets.
• Annotation and dataset management: Procedures for data labelling, preprocessing for annotation, training/testing splits and augmentation to support robust model training.
• Model selection and deep learning: Considerations for network architecture, input patching and use of deep learning techniques (informative annex).
• Evaluation and verification: Requirement to define evaluation metrics and procedures to quantify segmentation performance and to perform verification before 3D model generation.
• Post‑processing and deployment: Steps for refining segmentation outputs and integrating models into medical 3D printing workflows.
• Informative annexes: Practical guidance on CT scanning conditions, orbital bone segmentation challenges, deep learning techniques, and overall performance considerations.

Applications and users

ISO/IEC 3532-2:2024 targets organizations and professionals involved in medical 3D modelling and additive manufacturing, including:

• Medical device and surgical implant manufacturers
• Hospitals and clinical teams (radiologists, surgeons involved in preoperative planning, cranio‑maxillofacial specialists)
• Medical imaging and 3D printing software vendors
• Biomedical engineers, clinical engineers and AI developers working on segmentation models
• Regulatory and quality teams seeking standardized validation and traceability for patient‑specific devices

Practical applications include surgical planning, patient‑specific implant design (e.g., orbital wall reconstruction), anatomical modelling for education, and clinical workflow integration for medical 3D printing.

Related standards

• ISO/ASTM 52950 - Additive manufacturing - General principles - Overview of data processing
• ISO 15708‑1 - Non‑destructive testing - Radiation methods for computed tomography - Terminology
• ISO/IEC 2382 - Information technology - Vocabulary

Keywords: ISO/IEC 3532-2:2024, medical image segmentation, bone segmentation, CT segmentation, MR segmentation, 3D printing, medical 3D modelling, deep learning, annotation, preprocessing, segmentation evaluation.

Standards Content (Sample)

International
Standard
ISO/IEC 3532-2
First edition
Information technology — Medical
2024-02
image-based modelling for 3D
printing —
Part 2:
Segmentation
Technologies de l'information — Modélisation médicale à base
d'images pour l'impression 3D —
Partie 2: Segmentation
Reference number
© ISO/IEC 2024
All rights reserved. Unless otherwise specified, or required in the context of its implementation, no part of this publication may
be reproduced or utilized otherwise in any form or by any means, electronic or mechanical, including photocopying, or posting on
the internet or an intranet, without prior written permission. Permission can be requested from either ISO at the address below
or ISO’s member body in the country of the requester.
ISO copyright office
CP 401 • Ch. de Blandonnet 8
CH-1214 Vernier, Geneva
Phone: [phone]
Email: [e-mail]
Website: www.iso.org
Published in Switzerland
© ISO/IEC 2024 – All rights reserved
ii
Contents Page
Foreword .v
Introduction .vi
1 Scope . 1
2 Normative references . 1
3 Terms and definitions . 1
4 Abbreviated terms . 3
5 Objective of segmentation . 3
5.1 Background .3
5.2 Types of segmentation methods .4
6 Overall segmentation process . 4
6.1 General .4
6.2 Step 1: data preparation .5
6.3 Step 2: preprocessing for segmentation .5
6.4 Step 3: annotation .5
6.5 Step 4: selection of segmentation network model .5
6.6 Step 5: performance evaluation .5
6.7 Step 6: model deployment and running .6
6.8 Step 7: post-processing for segmentation .6
7 Data preparation . 6
7.1 General .6
7.2 Medical image .6
7.2.1 General .6
7.2.2 CT scan .6
7.2.3 MR image .6
7.3 Preparation steps .7
7.3.1 General .7
7.3.2 Image acquisition .7
7.3.3 Image reconstruction .7
8 Preprocessing for segmentation . 7
8.1 General .7
8.2 Intensity normalization .8
8.3 Spacing normalization .8
9 Annotation . 9
9.1 Data labelling .9
9.2 Preprocessing for annotation .9
9.3 Dataset management (training and testing) .10
9.4 Augmentation .10
10 Selection of network model . 10
10.1 General .10
10.2 Input patch .11
11 Evaluation .11
11.1 General .11
11.2 Evaluation metrics . 12
11.3 Evaluation procedure . 13
12 Deployment and running .13
13 Post-processing for segmentation . 14
Annex A (informative) CT scanning conditions for orbital bone segmentation .15
Annex B (informative) Characteristics of orbital bone segmentation from CT .16

© ISO/IEC 2024 – All rights reserved
iii
Annex C (informative) Deep learning techniques .18
Annex D (informative) Considerations for overall segmentation performance . 19
Bibliography .24

© ISO/IEC 2024 – All rights reserved
iv
Foreword
ISO (the International Organization for Standardization) and IEC (the International Electrotechnical
Commission) form the specialized system for worldwide standardization. National bodies that are
members of ISO or IEC participate in the development of International Standards through technical
committees established by the respective organization to deal with particular fields of technical activity.
ISO and IEC technical committees collaborate in fields of mutual interest. Other international organizations,
governmental and non-governmental, in liaison with ISO and IEC, also take part in the work.
The procedures used to develop this document and those intended for its further maintenance are described
in the ISO/IEC Directives, Part 1. In particular, the different approval criteria needed for the different types
of document should be noted. This document was drafted in accordance with the editorial rules of the ISO/
IEC Directives, Part 2 (see www.iso.org/directives or www.iec.ch/members_experts/refdocs).
ISO and IEC draw attention to the possibility that the implementation of this document may involve the
use of (a) patent(s). ISO and IEC take no position concerning the evidence, validity or applicability of any
claimed patent rights in respect thereof. As of the date of publication of this document, ISO and IEC had not
received notice of (a) patent(s) which may be required to implement this document. However, implementers
are cautioned that this may not represent the latest information, which may be obtained from the patent
database available at www.iso.org/patents and https://patents.iec.ch. ISO and IEC shall not be held
responsible for identifying any or all such patent rights.
Any trade name used in this document is information given for the convenience of users and does not
constitute an endorsement.
For an explanation of the voluntary nature of standards, the meaning of ISO specific terms and expressions
related to conformity assessment, as well as information about ISO's adherence to the World Trade
Organization (WTO) principles in the Technical Barriers to Trade (TBT) see www.iso.org/iso/foreword.html.
In the IEC, see www.iec.ch/understanding-standards.
This document was prepared by Joint Technical Committee ISO/IEC JTC 1, Information technology.
A list of all parts in the ISO/IEC 3532 series can be found on the ISO and IEC websites.
Any feedback or questions on this document should be directed to the user’s national standards
body. A complete listing of these bodies can be found at www.iso.org/members.html and
www.iec.ch/national-committees.

© ISO/IEC 2024 – All rights reserved
v
Introduction
This document was developed in response to the need for customization of 3D printing technology in the
medical industry through the use of information and communication technology (ICT).
There are many points where the existing standards for additive manufacturing (AM) do not match the
requirements of the medical industry. From medical images to 3D printing, medical device development is
quite a complex journey with complicated management of multiple pieces of software.
With the emerging market for medical 3D printed parts, there are many points requiring standardization.
There is currently no standardized process for the creation of protocols and validation procedures to ensure
that medical imaging data can be consistently and accurately transformed into a 3D printed object.
For medical 3D printing, segmentation techniques should be optimized and combined according to the
characteristics of the medical images and corresponding body parts to get an optimal 3D model.
In particular, during medical image segmentation, identification of the pixels of organs or lesions from raw
data such as computed tomography (CT) or magnetic resonance (MR) images, is one of the most challenging
analysis tasks.
For example, segmentation of the orbital bone is necessary for orbital wall reconstruction in cranio-
maxillofacial surgery to support the eye globe position and restore the volume and shape of the orbit.
However, orbital bone segmentation is challenging as the orbital bone is composed of cortical bone with a
high intensity value, and trabecular and thin bone with low intensity values, similar to soft tissue.
The human bone is delineated and extracted by segmentation techniques, and a 3D skeletal model is built
from this segmentation. The minimization of errors during segmentation of relevant body parts of interest
is critical. As there are several known critical issues for this segmentation, a verification process is made
before proceeding.
Not only single segmentation techniques but also combinations of those techniques should be adopted
for accurate extraction of a target body part. However, this process depends heavily on the operator. For
minimization of errors during this job, operators should know which segmentation technique is most used
in their imaging software and possess the necessary skills for that technique.
Thresholding techniques which are provided by a default Hounsfield unit (HU) range do not completely
[1]
recover true bony structure. An operator should typically adjust the extent of the segmentation manually.
The problem is usually under-segmentation. However, over-segmentation will also be problematic for
further designing processes, especially for surgical implants. Various techniques have been suggested to
[2]
reduce human error and improve performance and consistency for segmentation issues.
This document proposes a standardized process for the optimization of segmentation.

© ISO/IEC 2024 – All rights reserved
vi
International Standard ISO/IEC 3532-2:2024(en)
Information technology — Medical image-based modelling
for 3D printing —
Part 2:
Segmentation
1 Scope
This document provides an overview of the segmentation process for medical image-based modelling of
human bone. This document specifies a standardized process to improve the performance of human bone
segmentation.
This document is also applicable to medical 3D printing systems that include medical 3D modelling
capabilities.
2 Normative references
The following documents are referred to in the text in such a way that some or all of their content constitutes
requirements of this document. For dated references, only the edition cited applies. For undated references,
the latest edition of the referenced document (including any amendments) applies.
ISO 15708-1, Non-destructive testing — Radiation methods for computed tomography — Part 1: Terminology
ISO/IEC 2382, Information technology — Vocabulary
ISO/ASTM 52950, Additive manufacturing — General principles — Overview of data processing
3 Terms and definitions
For the purposes of this document, the terms and definitions given in ISO/ASTM 52950, ISO 15708-1,
ISO/IEC 2382 and the following apply.
ISO and IEC maintain terminology databases for use in standardization at the following addresses:
— ISO Online browsing platform: available at https:// www .iso .org/ obp
— IEC Electropedia: available at https:// www .electropedia .org/
3.1
image acquisition
scanning of the structure of interest using computed tomography (CT), magnetic resonance (MR) imaging or
other three-dimensional imaging technology
3.2
image annotation
process of attaching labels to an image
3.3
label
classifying phrase or name applied to a target

© ISO/IEC 2024 – All rights reserved
3.4
learning
process by which a biological or an automatic system gains knowledge or skills that it
may use to improve its performance
[SOURCE: ISO/IEC 2382:2015, 2122966, modified — Notes to entry have been removed.]
3.5
segmentation
process of separating the objects of interest from their surroundings
Note 1 to entry: Segmentation can be applicable to 2D, 3D, raster or vector data.
3.6
ground-truth label
correct answer of the training set for segmentation based on supervised learning
3.7
region of interest
ROI
specified boundary as defined in the image
3.8
machine learning
ML
process of optimizing model parameters through computational techniques, such that the model's behaviour
reflects the data or experience
[SOURCE: ISO/IEC 22989:2022, 3.3.5]
3.9
labelled data
group of data that have been tagged with one or more labels
3.10
hyperparameter
characteristic of a machine learning algorithm that affects its learning process
Note 1 to entry: Hyperparameters are selected prior to training and can be used in processes to help estimate model
parameters.
Note 2 to entry: Examples of hyperparameters include number of network layers, width of each layer, type of activation
function, optimization method, learning rate for neural networks; the choice of kernel function in a support vector
machine; number of leaves or depth of a tree; the K for K-means clustering; the maximum number of iterations of the
expectation maximization algorithm; the number of Gaussians in a Gaussian mixture.
[SOURCE: ISO/IEC 22989:2022, 3.3.4]
3.11
medical image
type of images generated by medical imaging devices

© ISO/IEC 2024 – All rights reserved
4 Abbreviated terms
AI artificial intelligence
AIM annotation and image markup
CAD computer-aided diagnosis
CCL connected component labelling
CNN convolutional neural network
CRF conditional random field
CT computed tomography
DICOM digital imaging and communications in medicine
DL deep learning
DSC dice similarity coefficient
FCN fully convolutional network
FOV field of view
FN false negative
FP false positive
HU Hounsfield unit
IoU intersection over union
MIoU mean intersection over union
ML machine learning
MR magnetic resonance
MRI magnetic resonance imaging
NIfTI neuroimaging informatics technology initiative
PET positron emission tomography
SPECT single-photon emission computerized tomography
TN true negative
TP true positive
US ultrasonography
5 Objective of segmentation
5.1 Background
The purpose of segmentation is to extract a specific region or organ from a patient's CT/MR medical image
and use it to create a 3D model.

© ISO/IEC 2024 – All rights reserved
Segmentation is the process of partitioning an image into different meaningful segments. For medical
images, segmentation techniques should be optimized and combined according to the characteristics of
image acquisition modalities and body parts to get an ideal 3D visualization. The human bone is delineated
and extracted by segmentation techniques, and a 3D skeletal model is built from this segmentation. The
minimization of errors during segmentation of relevant anatomy is critical.
Modelling for medical 3D printing requires optimized segmentation to provide better overlaying and
matching processes for human tissues and organs. However, most of the commercially available image
software cannot segment human bone effectively.
For improved medical image based 3D modelling, formalization and standardization of these procedures is
required.
5.2 Types of segmentation methods
[18][19][22][23]
Several methods have been investigated that segment human bone from CT images.
Semi-automatic segmentation: Thresholding is the simplest method of image segmentation, and
thresholding can be used to create binary images. This method replaces each pixel in an image with an
object pixel if the pixel intensity is greater than a specific human bone threshold value, or replaces it with a
background pixel if the pixel intensity is less than a specific human bone threshold value.
Deformable model-based segmenation: Deformable models are curves or surfaces for segmentation in
the image domain, which deform under the influence of internal and external forces to delineate the object
boundaries. The internal forces are defined such that they preserve the shape smoothness of the model,
while the external forces are defined by the image features to drive the model toward the desired region
boundary. By constraining extracted boundaries to be smooth and incorporating other prior information
[35][36]
about the shape, deformable models offer robustness to both image noise and boundary gaps.
However, there is a limit due to the difficulty of segmentation at the weak object boundary of a thin bone
with a low intensity value similar to soft tissue.
CNN-based segmentation: The FCN is a network that does not contain any dense layers. Instead, it contains
1x1 convolutions that perform the task of fully connected layers. The U-Net network, which is commonly
used for medical image segmentation, is based on a fully convolutional network and consists of a contracting
path and an expansive path, which gives it the U-shaped architecture.
6 Overall segmentation process
6.1 General
The overall segmentation process consists of seven steps in total, as described in 6.2 to 6.8. Figure 1 shows
the overall process flow of segmentation, where the numbers in parenthesis refer to clauses of this document.

© ISO/IEC 2024 – All rights reserved
Figure 1 — Overall process flow of segmentation
The software developer should implement the optimized segmentation process for medical image-based
modelling.
The considerations of the optimized segmentation process for medical image-based modelling should be
referenced from this document.
6.2 Step 1: data preparation
The objective of the data preparation stage is to transform the raw data so that the segmentation algorithm
can be applied. Detailed information is provided in Clause 7.
6.3 Step 2: preprocessing for segmentation
The objective of the preprocessing for segmentation stage is to normalize the data quality and induce
consistent segmentation results. This stage is an important step that affects the ability of the network model
to learn. Detailed information is provided in Clause 8.
6.4 Step 3: annotation
The objective of the annotation stage is to make the labelled training data fit for the learning of ML/DL-
based segmentation network models. Detailed information is provided in Clause 9.
6.5 Step 4: selection of segmentation network model
The objective of the segmentation network model selection stage is to select the optimal segmentation
network model. Detailed information is provided in Clause 10.
6.6 Step 5: performance evaluation
The objective of the performance evaluation stage is to calculate the agreement between the result of
applying the segmentation technique and the ground-truth label. Detailed information is provided in
Clause 11.
© ISO/IEC 2024 – All rights reserved
6.7 Step 6: model deployment and running
The objective of the deployment and running stage is to apply the optimally trained deep learning network
model to the 3D printing modelling system for achieving the maximized segmentation performance in a real
world environment. Detailed information is provided in Clause 12.
6.8 Step 7: post-processing for segmentation
The objective of the post-processing for segmentation stage is to refine the incorrect region after applying
the segmentation method. Detailed information is provided in Clause 13.
7 Data preparation
7.1 General
The objective of the data preparation stage is to transform the raw data so that the segmentation algorithm
can be applied.
7.2 Medical image
7.2.1 General
In the image acquisition phase, medical images are produced from devices, such as CT, MRI, PET/SPECT, US,
and optical scanners. Normally, they are reconstructed from “raw” or source data produced by detectors.
Generated medical image data from various devices should be stored in a standardized format (e.g. DICOM)
for medical image processing.
7.2.2 CT scan
Computed tomography (CT) is a medical imaging technique that uses X-rays to generate 2D slice images of
the body. These 2D slice images are typically stored in the DICOM file format. The intensity range of each
pixel is [−1024, 3072], because most CT images use 12 storage bits per pixel. Each pixel has its own value
according to the degree of transmitted radiation that passes through the body.
A CT scanner takes multiple radiographic projections and then uses an image reconstruction technique to
generate a series of 2D image slices covering a specific portion of the human body. The resolution of the
CT image is mainly dependent on pixel spacing and number of pixels, and partially dependent on the slice
thickness of the 2D images.
For obtaining better quality medical images, standardized CT scanning parameter conditions should be used
when scanning human bone. For example, in the case of the orbital bone, the minimum recommendation of
CT scanning conditions is defined in Annex A.
7.2.3 MR image
Magnetic resonance imaging (MRI) equipment is medical electrical equipment which is generally intended
for in vivo magnetic resonance examination of a patient. MRI is a medical imaging technique to form pictures
of the anatomy and physiological processes of the body. MRI scanners use strong magnetic fields, magnetic
field gradients and radio waves to create images of internal patient anatomy.
Standardized MRI scanning conditions or MRI scanning profiles should be considered.

© ISO/IEC 2024 – All rights reserved
7.3 Preparation steps
7.3.1 General
Since the data preparation stage consists of two steps, image acquisition and image reconstruction, the
essential considerations for each step should be reflected.
7.3.2 Image acquisition
The image acquisition step can be defined as the action of acquiring a set of image data from a hardware
source.
In order to create a good quality medical image-based 3D model for 3D printing, the following should be
considered during the image acquisition step.
1) High-quality medical imaging equipment should be used as much as possible and images should be
stored in the highest image quality.
2) Low-quality medical imaging equipment and protocols should not be used.
3) The acquisition of good quality medical images enables the production of good results.
4) Standardized image acquisition protocols should be used.
5) The image enhancement function provided by the device or manufacturer should be avoided as far as
possible.
NOTE Specific CT scanning protocols are recommended for delicate or complex structures, such as the orbital
wall, in order to optimize visualization of the entire structure. Annex A describes the CT scanning conditions for
orbital bone segmentation as an example of a specialized CT scanning protocol. See also Annex B.
7.3.3 Image reconstruction
The image reconstruction step can be defined as the mathematical process that generates composite images
from signals (or raw data) obtained during the image acquisition step.
The following points should be considered during the image reconstruction phase.
1) Image reconstruction can affect image quality.
2) There are many differences in the type and performance of CT image reconstruction kernels provided
by equipment manufacturers which should be considered.
NOTE The CT image reconstruction kernel, also known as a convolution algorithm, refers to the process used
to modify the frequency contents of projection data prior to back projection during image reconstruction in a CT
scanner. This process corrects the image by reducing blurring. The kernel affects the appearance of image structures
by sharpening the image. Different kernels have been developed for specific anatomical applications including soft
[5]
tissue (standard kernel) and bone (bone kernel).
8 Preprocessing for segmentation
8.1 General
The objective of the preprocessing stage is to normalize the image quality and induce consistent
segmentation results.
The following should be considered during the preprocessing step.
1) The potential sequence of preprocessing steps may be considered as: denoising, interpolation,
registration, organ windowing, followed by normalization, and potentially zero-padding to improve the
performance of segmentation.
© ISO/IEC 2024 – All rights reserved
2) Inconsistent intensity range and pixel spacing of an image can have a significant impact on the
performance of the segmentation method.
3) The normalization method may be required to ensure that the image training and testing data has a
consistent intensity range and pixel spacing.
8.2 Intensity normalization
This subclause specifically refers to images acquired by CT as an example. Intensity normalization is also
applicable, in principle, to images acquired using different technologies (see Figure 2).
Most CT images use 12 storage bits per pixel, but images used in DL approaches typically use 8 storage bits
per pixel. Considering the intensity range of the human bone, intensity normalization is performed so that
the HU intensity range [−200, 300] in the 12-bit CT image is transformed into the intensity range [0, 255], as
shown in Formula (1):
II−
max,n min,n
II=−()I +I (1)
min min,n
new
II−
maxmin
where
I is the pixel intensity value of the CT image;
I is the minimum intensity value of the orbit bones;
min
I is the maximum intensity value of the orbit bones;
max
I is the minimum intensity value of the new image;
min,n
I is the maximum intensity value of the new image.
max,n
a) Original 12-bit CT image (level: 400 HU, width: b) Transformed 8-bit CT image after intensity
1000) normalization
Figure 2 — Intensity normalization in head-and-neck CT image
8.3 Spacing normalization
This subclause specifically refers to images acquired by CT as an example. Spacing normalization is
applicable, in principle, also to images acquired using different technologies (see Figure 3).

© ISO/IEC 2024 – All rights reserved
CT images can have different pixel spacing depending on patients. The difference between the pixel spacing
means that the interval considered by one pixel is different. To normalize the pixel spacing, the pixel spacing
normalization is performed using the maximum or minimum spacing of the dataset
a) 8-bit CT image with original pixel spacing b) Transformed 8-bit CT image after spacing nor-
(0.49 mm) malization (0.53 mm)
Figure 3 — Spacing normalization in head-and-neck CT images
9 Annotation
9.1 Data labelling
Labelled data is used as input data in the training stage to find the optimal parameters of the segmentation
network.
Machine learning algorithms typically fall within the domain of supervised artificial intelligence and are
designed to "learn" from annotated data. Machine learning models require large, diverse training datasets
for optimal model convergence. This means that before an ML algorithm can be trained and tested, the
[31]
ground truth (annotation) needs to be defined and linked to the image.
For training of segmentation network model preparation, the following should be considered during the
labelling data step.
1) Well-annotated datasets should be prepared. Well-annotated datasets are crucial to training accurate,
generalizeable ML algorithm models.
2) There are many types of image annotation (e.g. closed curve, curve, ellipse, freehand, line, point, pointer,
polygon, polyline, rectangle, text, text pointer, etc.). Any type of image annotation should be used in an
interoperable way.
3) Annotation (or labelling) format should support annotation interoperability, such as AIM (annotation
and image markup), NIfTI-1 data format or DICOM.
4) Consistent pixel values should be used for labelling. For example, the pixel value of the segmented
human bone region can be set as 1 and the pixel value of the other region can be set as 0.
9.2 Preprocessing for annotation
The objective of the preprocessing stage is to normalize the data quality and to enhance the training ability
of the segmentation network model.
The considerations for this stage are nearly the same as those described in Clause 8.

© ISO/IEC 2024 – All rights reserved
9.3 Dataset management (training and testing)
The objective of the data management stage is to effectively manage all data (training, validation, test)
needed to optimize the actual performance of the segmentation network model.
Data management procedures should be conducted in the following order.
a) Collect the raw medical image data.
b) Apply data labelling.
c) Select a data sampling strategy for training/validation/test.
d) Split the datasets for training/validation/test.
The following should be considered during the data management step.
1) Collect as much high-quality data as possible. Performance increases logarithmically based on volume
[32]
of training data.
2) A well-designed dataset should be prepared which increases the quality of the resulting segmentation
network model.
3) Even the most sophisticated segmentation network model cannot be trained with poor quality data.
4) A management process should be used for data quality.
5) Data should be assigned separately for training, validation and test purposes, as follows.
— Training dataset: The set of data used to fit the model.
— Validation dataset: The set of data used to provide an unbiased evaluation of a trained model fitting
on the training dataset while tuning model hyperparameters.
— Test dataset: The set of data used to provide an unbiased evaluation of a final trained model fitting
on the training dataset.
9.4 Augmentation
The objective of the augmentation stage is to generate the augmented dataset which increases the amount
of data by adding slightly modified copies of already existing data or newly created synthetic data from
existing data.
[37]
There are different types of data augmentation techniques. Increasing the dataset size by data warping
or oversampling is generally used. Data warping augmentations transform existing images such that their
labels are preserved, which includes augmentations such as geometric and color transformations, random
erasing, adversarial training, and neural style transfer. Oversampling augmentations create synthetic
instances and add them to the training set, which includes image mixing, feature space augmentations, and
generative adversarial networks.
Data augmentation is useful as it improves the performance and outcomes of machine learning models
by forming new and different examples for datasets. Data augmentation can provide benefits such as
solving small dataset problems, preventing data scarcity, reducing data overfitting, and increasing the
generalization ability.
10 Selection of network model
10.1 General
The objective of the network model selection stage is to find and select the optimal network model and
model parameters for segmentation.

© ISO/IEC 2024 – All rights reserved
The broad success of DL has prompted the development of new image segmentation approaches leveraging
ML/DL models. More than 100 DL-based image segmentation models have been developed (see Annex C)
which have differences in network architecture selection, training data, loss function, training strategy and
evaluation metrics.
The overall steps for finding and selecting the best segmentation model should proceed in the following
order.
1) Divide the available data into training, validation and test dataset.
2) Select a model and set the training hyperparameters.
3) Train the model using the training set.
4) Evaluate the model using the validation dataset.
5) Repeat steps 2) through 4) using different model and training parameters.
6) Select the best model and train it using data from the training and validation dataset.
7) Assess this final model using the test dataset.
8) If external validation is required, assess the final model using the external validation dataset or open
dataset
The evaluation method for finding and selecting the best segmentation model shall use the procedure and
metrics defined in Clause 11.
10.2 Input patch
A 2D image patch is usually a .jpg image file format with 8 storage bits per pixel. The input patch consists of
an image patch and a mask patch. The image patch is used in the training and test stages. The mask patch is
used only in the training stage (see Figure 4).
a) Image patch for CT images b) Mask patch for labelling data
Figure 4 — Input patch
11 Evaluation
11.1 General
The objective of the evaluation stage is to assess the differences between the results of applying the trained
segmentation model and the ground-truth labels.

© ISO/IEC 2024 – All rights reserved
Comparing results to evaluate the quality of segmentation is an essential part of finding the best
segmentation model and algorithm. Annex D provides an example of how to consider overall segmentation
performance in the case of the orbital bone.
There are different quality aspects in medical image segmentation based on which types of segmentation
errors can be defined. Evaluation metrics for segmentation can be selected to indicate these errors,
[32][33]
depending on the data and the target.
Several classification metrics are derived from the following elements:
— true positive, T , is the number of pixels correctly identified as positive;
P
— true negative, T , is the number of pixels correctly identified as negative;
N
— false positive, F , is the number of pixels incorrectly identified as positive;
P
— false negative, F , is the number of pixels incorrectly identified as negative.
N
11.2 Evaluation metrics
Software developers should select the evaluation metrics that can best evaluate the quality and performance
of their segmentation models, and use those evaluation metrics to select the best segmentation model.
Performance measurements for segmenting the human bone are calculated using sensitivity, specificity,
accuracy, DSC and IoU. T , F , T and F refer to the detected number of pixels which are true positives, false
P P N N
positives, true negatives and false negatives, respectively.
Sensitivity is the ability to determine the labelled human bone areas correctly. Sensitivity measures the
portion of true positive in the labelled human bone area of the ground truth, and is defined as:
T
P
s=
TF+
PN
where s is the calculated sensitivity value.
Specificity is the ability to determine the background correctly. Specificity measures the portion of true
negative in the background of ground truth, and is defined as:
T
N
S =
TF+
NP
where S is the calculated specificity value.
Accuracy is the ability to differentiate the labelled human bone areas and background correctly. Accuracy
measures the proportion of true positive and true negative in all ROI, and is defined as:
()TT+
PN
A=
TF++TF+
()
PP NN
where A is the calculated accuracy value.
Dice similarity coefficient (DSC) is the most widely used metric when validating medical volume
segmentation, and is defined as:
2T
P
D=
2TF++F
PPN
where D is the calculated dice similarity coefficient value.

© ISO/IEC 2024 – All rights reserved
Intersection over union (IoU) is the Jaccard index and is defined as:
T
P
i =
TF+ ++TF −T
() ()
PP PN P
where i is the calculated intersection over union value.
Additional evaluation metrics can exist which are not described here.
NOTE 1 Evaluation metrics are used in the evaluation procedure.
NOTE 2 ISO/IEC TS 4213 provides evaluation metrics for classification tasks.
[33][34][38][39]
NOTE 3 Evaluation metrics are also often used as the loss function for deep learning segmentation.
11.3 Evaluation procedure
The evaluation procedure for assessing the performance of segmentation techniques should be performed
in three steps using the evaluation metrics.
First, the ground-truth label should be obtained from human experts.
Second, the segmentation result should be obtained by applying the segmentation technique.
Third, the discrepancy between the segmentation result and the ground-truth label should be measured.
The method of measuring discrepancies should be counted as the difference between a segmented result
and a correctly segmented image (ground truth) using evaluation m
...