Інтерактивна система підтримання прийняття рішень для сегментації раку легенів
This paper studies Clinical Intelligent Decision Support Systems (CIDSSs) for lung cancer segmentation, which are based on deep neural nets. A new interactive CIDSS is proposed and compared with previous approaches. Additionally, the purpose uncertainty problem in building interactive systems is dis...
Saved in:
| Date: | 2024 |
|---|---|
| Main Author: | |
| Format: | Article |
| Language: | English |
| Published: |
The National Technical University of Ukraine "Igor Sikorsky Kyiv Polytechnic Institute"
2024
|
| Subjects: | |
| Online Access: | https://journal.iasa.kpi.ua/article/view/290923 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Journal Title: | System research and information technologies |
| Download file: |  |
Institution
System research and information technologies| _version_ | 1866302939903033344 |
|---|---|
| author | Sydorskyi, Volodymyr |
| author_facet | Sydorskyi, Volodymyr |
| author_sort | Sydorskyi, Volodymyr |
| baseUrl_str | http://journal.iasa.kpi.ua/oai |
| collection | OJS |
| datestamp_date | 2024-08-11T01:12:49Z |
| description | This paper studies Clinical Intelligent Decision Support Systems (CIDSSs) for lung cancer segmentation, which are based on deep neural nets. A new interactive CIDSS is proposed and compared with previous approaches. Additionally, the purpose uncertainty problem in building interactive systems is discussed, and criteria for measuring both quality and amount of user feedback are proposed. In order to automate system evaluation, a new algorithm was used to simulate expert feedback. The proposed interactive CIDSS outperforms previous approaches (both interactive and noninteractive) on the task of lung lesion segmentation. This approach looks promising both in terms of quality and expert user experience. At the same time, this paper discusses a bunch of possible modifications that can be done to improve both evaluation criteria and proposed CIDSS in future works. |
| doi_str_mv | 10.20535/SRIT.2308-8893.2024.2.05 |
| first_indexed | 2025-07-17T10:28:24Z |
| format | Article |
| fulltext |
V. Sydorskyi, 2024
68 ISSN 1681–6048 System Research & Information Technologies, 2024, № 2
UDC 004.424.4
DOI: 10.20535/SRIT.2308-8893.2024.2.05
INTERACTIVE DECISION SUPPORT SYSTEM FOR LUNG
CANCER SEGMENTATION
V. SYDORSKYI
Abstract. This paper studies Clinical Intelligent Decision Support Systems
(CIDSSs) for lung cancer segmentation, which are based on deep neural nets. A new
interactive CIDSS is proposed and compared with previous approaches. Addition-
ally, the purpose uncertainty problem in building interactive systems is discussed,
and criteria for measuring both quality and amount of user feedback are proposed. In
order to automate system evaluation, a new algorithm was used to simulate expert
feedback. The proposed interactive CIDSS outperforms previous approaches (both
interactive and noninteractive) on the task of lung lesion segmentation. This ap-
proach looks promising both in terms of quality and expert user experience. At the
same time, this paper discusses a bunch of possible modifications that can be done
to improve both evaluation criteria and proposed CIDSS in future works.
Keywords: clinical decision support systems, deep learning, open system, interac-
tive segmentation.
INTRODUCTION
Decision support systems (DSS) stand out as instrumental across various domains
of human activities [1–3]. These systems combine the advantages of big data, sta-
tistical models, big informational systems, classical machine learning, and deep
learning technologies. Of particular academic and practical interest are the emerg-
ing Intelligent DSS. Characterized by their reliance on neural networks, these sys-
tems promise enhanced analytical depth and precision, setting the stage for trans-
formative applications across diverse sectors.
A significant portion of recent research [4] has been dedicated to the applica-
tion of machine learning and deep learning in medical imaging. These studies
primarily focus on architecting neural networks and formulating optimization
policies to improve diagnostic accuracy. However, it is important to consider deep
learning models as part of DSS systems and sub-systems in medical applications.
Being a part of IDSS, they can benefit from advances in computer science, system
analysis, and decision-making in scientific and practical spheres. IDDS, based on
neural nets, can provide accurate medical insights, a user-friendly interface, and
an interactive and adaptive mechanism for decision-making.
To address the problem of building IDSS based on deep learning ap-
proaches, using expert feedback to improve the initial results of the proposed sys-
tem, the task of lung cancer segmentation is considered. The main contributions
of this paper are:
Integration of a segmentation neural net into an interactive intelligent de-
cision support system.
Interactive decision support system for lung cancer segmentation
Системні дослідження та інформаційні технології, 2024, № 2 69
Adaptation of the previous interactive segmentation approach [5] to the
task of lung cancer segmentation.
Improvement of the previous approach by the usage of two types of seg-
mentation neural nets.
Formulate a purpose uncertainty problem [6] in the construction of such
a DSS and propose a set of criteria for DSS assessment, including a special algo-
rithm for expert feedback simulation.
Also, it is important to mention that the proposed system can be adapted for
other types of cancer diagnosis. Such adaptions will be considered in future
works.
RELATED WORK
The application of clinical decision support systems (CDSS) was proposed in re-
cent studies in order to help with the diagnosis and treatment of oncology diseases
[7], modifications in order to adapt to cancer treatment in developing countries
[8], and specialized DSS, for example for brain tumors treatment [9]. All these
works are focused on building complete decision support systems. Still, at the
same time, they focus less on particular intelligent subsystems, which hugely
benefit the precision of the overall system.
The latest research in computer science proposes a wide variety of deep
learning algorithms [10], which can be utilized in order to improve the quality of
CDSS. Starting from classical U-Net [11], FPN [12], and DeepLab [13] architec-
tures, which made a breakthrough in semantic image segmentation in the sphere
of medical imaging, ending with the latest Unet++ [14], UNet 3+ [15] and
UNETR [16] architectures, which rely on complex skip and residual connections
and on processing the whole image volume altogether (3D Nets). Additionally,
different Hybrid and Neuro-Fuzzy Networks were applied to the task of lung and
brain tumor segmentation and detection in combination with classical 2D, MIL,
and 3D approaches [17].
At the same time, interactive segmentation methods propose guiding deep
learning algorithms with user feedback. Segment Anything Model [18] proved
that neural nets can generalize to the segmentation of many different objects in
completely different scenarios. However, there is doubt that it can segment pre-
cise medical images, especially in the presence of very tiny and hard distinguish-
able objects, like lung cancer lesions. Other more specialized approaches were
proposed [19; 5] and showed success in medical image segmentation.
INTERACTIVE INTELLIGENT DECISION SUPPORT SYSTEM FOR LUNG
CANCER SEGMENTATION
This work considers modern deep learning algorithms to build an interactive intel-
ligent decision support system for lung cancer segmentation. The proposed ap-
proach operates on CT scans from different manufacturers and incorporates user
feedback to increase the segmentation quality. Finally, three possible architectures
of DSS are compared by several criteria.
V. Sydorskyi
ISSN 1681–6048 System Research & Information Technologies, 2024, № 2 70
First, let’s formalize the main problem that the segmentation system should
solve. This system receives a pre-processed CT scan — x , which is a 3D image,
where dimensions refer to the width, height, and depth of the image matrix:
slicesnkwidthjheightixx ijk _,...,0,,...,0 ,,...,0 },{ .
In the proposed approach, the Multi-Instance Learning (MIL) [17; 20] ap-
proach is used, where each instance refers to a slice of the scan. So, the basic deep
learning model — H should approximate mask pixel distribution for each slice
based on several scan slices:
,, ,)(
IJkykIJKKkK yxxyxH
},...,,{ },,...,1,0{}, ,...,1,0{ 21 lkkkKwidthJheightI ,
where K refers to some set of indexes that define slices, which are used for the
prediction of each slice mask, k refers to some fixed slice index, and
ky refers
to an approximated slice mask.
In order to make the system interactive, it is required to introduce additional
input to the network, which refers to the expert feedback on initially approxi-
mated slice masks. One of the ways to incorporate such feedback is to allow ex-
pert to select pixels that, he thinks, are misclassified. An expert can simply click
on pixels that he thinks should be assigned to foreground — lesion region or
background — non lesion region. Also, an expert can erase the whole slice mask.
Formalization of expert’s feedback:
Two sets of clicks are provided by the expert: pc _ — “positive” clicks,
which refer to the foreground ( }, ,{_ jipc },,...,1,0{ heighti
} ,...,1,0{ widthj ) and nc _ — “negative” clicks, which refer to the background
( } ,...,1,0{},,...,1,0{ },,{_ widthjheightijinc ).
Click mask — cm ( },{ ijcmcm , ,...,0 heighti , ,...,0 widthj
), 0[ ijcm ).
Algorithm for encoding clicks into click mask MC2
( _)_(2 OR _)_(2 ncmncMCpcmpcMC ).
New model — IH _ (
kKxk fyyncmpcmIH _),_,_(_ , ).
It is important to mention that IH _ takes the previous predicted mask
ky
on input, which allows to incorporate both information from the previous step and
expert feedback. The interactive procedure can be applied several times: redefine
ky :=
kfy _ and ask expert for new pc _ and nc _ . Such an iterative procedure
creates a certain trade-off: more iterations result in better segmentation masks
while taking more expert’s time. Also, to receive the initially predicted mask, two
approaches can be used: predicted masks from H or initialize
ky as zero mask
and ask an expert to define foreground areas without the initially predicted mask.
So there are three possible structures of segmentation DSS:
Interactive decision support system for lung cancer segmentation
Системні дослідження та інформаційні технології, 2024, № 2 71
1. Noninteractive (Fig. 1).
2. Interactive, without the initially predicted mask (Fig. 2).
3. Interactive with initially predicted mask, received from H (Fig. 3).
2. Non Interactive IDSS
Input and Process
Image
x
H
Expert
y_f
Fig. 1. Structure of noninteractive IDSS
Fig. 2. Structure of interactive IDSS, without the initially predicted mask
2. Interactive, without the initial predicted mask IDSS
Input and Process Image
x
H_I
Expert
c_n and
c_p C2M
cm_p and
cm_n
y_f
Fig. 3. Structure of interactive IDSS with initially predicted mask received from H
Input and Process Image
Expert
3. Interactive with Initial predicted mask, received from H, IDSS
x
H
c_n and
c_p C2M
y_f
H_I
cm_p and
cm_n
y_f
V. Sydorskyi
ISSN 1681–6048 System Research & Information Technologies, 2024, № 2 72
Although, it is important to mention that proposed segmentation DSSs are
open systems because they explicitly have information exchange in the form of
user feedback. While utilizing this user feedback, entropy in the system is de-
creasing. This can be easily tracked by the improvement of quality metrics, which
will be discussed in Experiments section.
Purpose uncertainty problem. One of the problems in building such an
IDSS is purpose uncertainty [6] because there is a trade-off between the amount
of user interaction and the quality of the final mask. In this task, different ways of
expert interaction require different amounts of effort:
– Making a positive click is the most challenging action because it requires
an understanding of precise lesion location, while the initial mask, produced by
DSS, can bias the expert.
– Making negative clicks requires less effort because, in most cases, finding
a healthy region is an easier task.
– Erasing the whole mask is the most straightforward action because it does
not require selecting a particular region.
Considering quality metrics, there is also some uncertainty because the pro-
posed systems will be tested on a series of CT scans, but it is important to track
both average quality results and a number of poorly segmented masks. In order to
handle this problem, median metric values and the whole metric distribution will
be considered in Evaluation section.
DETAILED NEURAL NET SETUP
First of all, K set for selecting input instances should be defined. In the proposed
approach kkkkkkK },2,1,,1,2{ refers to target slice.
So, the neural net will operate on the target slice and two previous and next
slices. Initial experimental results proved that using more slices does not improve
segmentation quality and only leads to optimization and inference overheads and
overfitting. To feed selected slices in regular 2D Convolutions, slices will be
stacked on channel dimension so the resulting input will have the following form:
},...,1,0{ },,...,1,0{,__ widtJheightIKIJxkx .
In this paper, classical Encoder-Decoder segmentation neural nets will be
used: EfficientNet B3 [21] encoder and Unet++ [14], DeepLabV3+ [22] decoders.
Different setups will be evaluated in Experiments section, and the best one will be
chosen according to the proposed criteria. After the final fully connected layer,
the sigmoid will constraint outputs of each pixel prediction to [0; 1] range, and
thresholding by 0.5 will be applied.
To design click-to-mask transformation, “Gaussian” masks will be used, as
proposed in [5] — equations:
)), ,(( },{ lijijc cjidddDM
l
;
0else _ if },{
))2/(( 22
radiusclipdecmcmCM ij
sigmad
ijijc
ij
l
;
cl
c
CMCM
l
,
Interactive decision support system for lung cancer segmentation
Системні дослідження та інформаційні технології, 2024, № 2 73
where
lcDM is a distance transform for each click lc ; d — euclidean distance;
lcCM — click mask for each click lc ; sigma — “Gaussian standard deviation”;
radiusclip _ parameter, which zeros out mask pixels, which are too far from the
center; CM final click mask for several clicks. In the proposed approach, sepa-
rate masks for positive and negative clicks are used. Finally, to feed clicks masks
and previously predicted masks to an interactive neural net IH _ , all inputs are
stacked channel-wise, so IH _ has three more additional input channels com-
pared to H .
DETAILED OPTIMIZATION SETUP
For training both H and IH_ Adam optimizer [23] and reducing the learning
rate with the OneCycleCosine learning rate scheduler, starting with 1e-3 and fin-
ishing with 1e-6, are used. Neural nets are trained for 191000 steps. Batch size is
set to 32, and a distributed data parallel [24] mechanism on two A100 GPUs is
used.
It is important to consider the data distribution of masked regions, which is
typical for most medical segmentation tasks. In most cases, regions with lesions
take only a very small partition of the whole image area, so there are many more
negative pixels than positive ones. This work proposes to sample slices with and
without lesions with equal probability to reduce class imbalance for the optimiza-
tion procedure.
In order to optimize both neural nets, the sum of BCE and Jaccard [25]
losses is used.
Also, a transfer learning mechanism is used, and the encoder is initialized
with weights received from a noisy student learning procedure [26].
For training noninteractive net ( H ) horizontal and vertical flip augmenta-
tions are used, and on the inference stage, the same test time augmentation [27] is
applied to improve prediction quality. For training interactive net ( IH _ ) any
augmentations are not used.
Another important part is the click-sampling strategy. This work mostly fol-
lows strategies proposed in [5] but with several modifications specific to the
medical imaging domain. Initially, there are not previously predicted masks, so
first clicks should be initialized by some “cold start” strategy. For positive clicks,
the center masses of each lesion are selected; if there are several lesions on one
slice, the lesion with the biggest area is picked. For negative clicks, random points
within 0d distance from the lesion are selected. Other negative click sampling
strategies from [5] were not considered because of task specifications:
Sampling from another object is not possible in this task because there is
only one type of object to segment.
Sampling from the target object border is also inefficient because the le-
sion area is pretty tiny, and its border may contain a big part of this area.
After sampling clicks, a click mask for the current interaction can be created
and saved to the click cache for the next interactions. Predicted masks for click
sampling are used only after the first epoch. Having previously predicted masks
V. Sydorskyi
ISSN 1681–6048 System Research & Information Technologies, 2024, № 2 74
and previous clicks, it is possible to use other sampling strategies. For selecting a
positive click, the algorithm uses false negative pixels connected regions and se-
lects the biggest area region and sample click, which is located on the largest dis-
tance both from the region border and previous clicks:
))),(min), ,(min(min(maxarg cprppl ppdppdc
ct
, (1)
where lc refers to click to select; p — all possible points; tp — points of the
region; cp — previous clicks from the cache. The algorithm follows the same
logic for negative clicks, but here, false positive masks are used as mislabeled
masks. To introduce diversity and ensure that the neural net can work only with a
few clicks, the algorithm randomly resets the click cache for each image with a
probability of 0.3. Also, the maximum number of positive and negative clicks is
constrained to 12 separately. It is done to reduce RAM overhead and ensure that
the neural netdoes not abuse the usage of lots of clicks.
EXPERIMENTAL RESULTS
In this section, the following points are described:
– Lung lesions dataset, which is used for training neural networks and testing
proposed DSSs;
– Metric that is used for evaluation and solution for purpose uncertainty issue;
– An algorithm for simulating feedback from an expert;
– Results of the proposed approach.
Dataset. To experiment with the proposed IDSS, the Lung Image Database
Consortium image collection (LIDC-IDRI) [28] is used, which consists of diag-
nostic and lung cancer screening thoracic computed tomography (CT) scans with
marked-up annotated lesions. In this research, only CT scans and marked-up an-
notated lesions are used.
The whole dataset contains 1018 scans from 1010 unique patients. So, in
most cases, there is one CT scan for one patient. The dataset annotation was ob-
tained from four independent experts. To get the final mask, the consensus of the
majority of annotators [29] was used. The examples of original and masked im-
ages are presented in Fig. 4.
All CT scans have 512512 height and width, but they all differ in the
number of slices. Slices distribution can be seen in Fig. 5.
Fig. 4. Examples of CT scans with annotation. First row — original scans. Second row
— annotation
Interactive decision support system for lung cancer segmentation
Системні дослідження та інформаційні технології, 2024, № 2 75
As mentioned before, the biggest problem in lesion segmentation is a severe
class imbalance — most of the pixels are background pixels. Moreover, most of
the slices do not contain lesions: the median number of slices in the scan is 203.5,
and the median number of slices with lesions is 40. Also, the area of lesions is
pretty small; the median of the relative area of the foreground area to the whole
scan area across all dataset images is 8.7683e-06. So, such class imbalance should
be considered during model optimization and evaluation.
Evaluation. Two main criteria to evaluate results will be considered: quality
and amount of expert feedback.
For quality evaluation, the Intersection over Union (IoU) [30] metric is used,
which is computed for each scan separately and averaged across all scans. If the
scan does not contain lesions and the predicted mask is empty — IoU is equal
to 1. This work also outlines the median, first ( 1Q ), and third ( 3Q ) quartiles of
IoU scores and overall distribution.
For the amount of expert feedback evaluation, three types of feedback, out-
lined in Purpose uncertainty problem section, are considered. The feedback score
equation (2) includes different weights for feedback types. These weights are pro-
posed based on personal perception of the process and can be tuned in future work.
ClicksPositiveofNumberScoreFeedback
85.0 * clicksNegativeofNumber
75.0 * erasementsMasksofNumber . (2)
This score should be minimized.
The dataset is split into train and test sets with stratification by relative area
of foreground area and grouping by patients (so scans of one patient are only in
the test or train set). After the split, there are 203 scans in the test set and 815
scans in the train set, so approximately 20% of the data is used for model assess-
ment.
Algorithm for simulating expert feedback. This algorithm is needed to test
our second and third types of IDSS. The proposed algorithm behaves as an
“ideal” expert. The algorithm makes one optimal click for each slice in each scan
Fig. 5. Distribution of the number of slices across the dataset
V. Sydorskyi
ISSN 1681–6048 System Research & Information Technologies, 2024, № 2 76
on each iteration. Such an approach is impossible in the real scenario, but profes-
sional radiologists should behave near the “ideal” behavior. This work proposes
possible improvements for future work:
Introduce the probability of making positive, negative clicks or mask
erased.
Introduce some probability distribution of click coordinate.
Condition both previous distributions on lesion size, number of ground
truth, and predicted lesions in the particular CT scan.
Reduce the probability of correct feedback with iteration number.
Let’s formalize optimal positive and negative clicks. First, false positive pix-
els connected regions for positive click and false negatives connected regions for
negative are computed. If there are no such regions for a particular scan or the
whole region consists only of its border (single-pixel line), the algorithm just
omits clicking. Otherwise, it selects the biggest mislabeled region using the tech-
nique proposed in equation (1).
If no feedback was done for a particular slice, a mask from the previous it-
eration or an empty mask if we have a “cold start” scenario is used.
Experiments. Both different systems and different decoders for neural net-
works H and _H I will be compared by evaluation metrics proposed in
Evaluation section
Quality metrics are outlined in Table 1.
T a b l e 1 . Quality metrics
System and Decoder Mean IoU Median IoU 1 Q IoU 3 Q IoU
System 1
Unet++ Decoder
0.4083 0.4385 0.1507 0.5993
System 1
DeepLabV3+ Decoder
0.3929 0.4104 0.1202 0.6103
System 2
Unet++ Decoder
0.7031 0.6904 0.6154 0.7729
System 2
DeepLabV3+ Decoder 0.6868 0.6679 0.5925 0.7587
System 3
Unet++ Decoder 0.7182 0.7039 0.6351 0.7945
System 3
DeepLabV3+ Decoder
0.6986 0.6833 0.6063 0.7844
DeepLabV3+ and Unet++ decoders are compared because the first one was
used in [5]. From the IoU score, it is obvious that Unet++ outperforms the Dee-
pLabV3+ decoder for all systems, which is logical because the model can benefit
from lots of skip connections in Unet++ architecture. Bootstrapped cross-entropy
[31] loss was also tested in this work, but the results were worse than the pro-
posed loss setup. Comparing different systems, the proposed System 3 outper-
forms all other systems by all quality metrics and with all proposed decoders. It is
essential to pay attention to IoUMean , and IoU1Q : we can see that the
interactive approach introduces more than 0.3 improvements in mean score and
a bit less than 0.5 improvement in 1Q score, which is a huge increase in quality.
Interactive decision support system for lung cancer segmentation
Системні дослідження та інформаційні технології, 2024, № 2 77
Next, we can observe IoU scores distribution for System 1 Unet++ Decoder,
System 2 Unet++ Decoder, and System 3 Unet++ Decoder in Fig. 6.
Obtained results show fewer scans segmented with very poor scores and
many more scans segmented with nearly ideal scores.
Visual results of segmentation can be seen in Figs. 7, 8.
According to visual results, the model correctly reacts to positive and nega-
tive clicks. Additionally, the model does not remove the correct segmentation if
negative clicks are placed near correctly segmented regions. However, there is
still space for improvement. As we can see from Fig. 7, the model is pretty con-
servative in segmenting additional areas after a positive click for some scans.
Feedback Score (equation (2)) is outlined only for interactive systems
(second and third) in Table 2.
T a b l e 2 . Number of all feedback types and resulting Feedback Score
System and Decoder
Number of
Positive
Clicks
Number
of Negative
Clicks
Number
of Masks
Erasements
Feedback
Score
System 2
Unet++ Decoder 3024 0 0 3024
System 2
DeepLabV3+ Decoder 3024 0 0 3024
System 3
Unet++ Decoder 1637 284 1384 2916.4
System 3
DeepLabV3+ Decoder 1730 248 1344 2948.8
1 2
3
Fig. 6. IoU scores distribution for Systems 1, 2, and 3 with Unet++ decoder
V. Sydorskyi
ISSN 1681–6048 System Research & Information Technologies, 2024, № 2 78
From Table 2: System 2 requires only positive clicks because there is no ini-
tial mask approximation, while System 3 requires correction of initial results, so it
has all types of feedback. Considering weighting coefficients from equation (2),
we see that System 3 again outperforms System 2 with all decoders.
Considering both quality and feedback criteria, we can conclude that System
3 with Unet++ decoder is a rational choice for the lung cancer segmentation task.
Finally, let’s investigate the systems’ performance on more interactive feed-
back iterations in Fig. 9.
From the received plots, System 2 outperforms System 3 in mean IoU only
from the third feedback iteration while still losing in feedback score. These results
are logical because System 2 always relies on expert feedback, while System 3
has initial mask approximation, which is not conditioned on feedback. On the
other hand, there is still an issue with the “ideal” behavior for users’ feedback
simulation, and this issue will be addressed in future works.
Fig. 7. Results of segmentation from Sytem 3 Decoder Unet++. First row — original
scans. Second row — annotation. Third row — initial masks from H . Fourth row —
positive clicks. Fifth row — corrected masks with IH _
Interactive decision support system for lung cancer segmentation
Системні дослідження та інформаційні технології, 2024, № 2 79
CONCLUSIONS
This paper has studied different IDSS for lung cancer segmentation, proposed
evaluation criteria, and the algorithm for users’ feedback simulation. Finally, a
Fig. 8. Results of segmentation from Sytem 3 Decoder Unet++. First row — original
scans. Second row — annotation. Third row — initial masks from H . Fourth row —
negative clicks. Fifth row — corrected masks with _H I
1 — System 2 Unet ++ Decoder
2 — System 2 Unet ++ Decoder
IoU depending on Feedback iteration number
Io
U
Iteration
1 — System 2 Unet ++ Decoder
2 — System 3 Unet ++ Decoder
1
2
Feedback score depending on Feedback iteration number
Fe
ed
ba
ck
s
co
re
1 — System 2 Unet ++ Decoder
2 — System 3 Unet ++ Decoder
Iteration
1
2
Fig. 9. Mean IoU and Feedback Score depending on feedback iteration number for
Systems 2 and 3 with Unet++ Decoder
V. Sydorskyi
ISSN 1681–6048 System Research & Information Technologies, 2024, № 2 80
system that outperforms previous approaches by all criteria is proposed. However,
it may have less quality increase with feedback iterations compared to previous
systems. This issue will be addressed in future works. It is important to mention
that the proposed system is a combination of 2 other systems, so it is possible that
there is a huge room for improvement. Another issue that requires further research
is the users’ feedback simulation algorithm. Its behavior is too “ideal,” and there
should be randomness conditioned on previous mask approximation results.
REFERENCES
1. R.T. Sutton, D. Pincock, D.C. Baumgart, D.C. Sadowski, R.N. Fedorak, and K.I. Kroeker,
“An overview of clinical decision support systems: benefits, risks, and strategies for suc-
cess,” NPJ Digit. Med., pp. 3–17, 2020 Feb 6. doi: 10.1038/s41746-020-0221-y.
2. E. Walling, C. Vaneeckhaute, “Developing successful environmental decision sup-
port systems: Challenges and best practices,” Journal of Environmental Manage-
ment, 264, 110513, 2020. doi: 10.1016/j.jenvman.2020.110513.
3. M. Rashidi, M. Ghodrat, B. Samali, and M. Mohammadi, “Decision Support Sys-
tems,” Management of Information Systems, 2018. doi: 10.5772/intechopen.79390.
4. N. Shahid, T. Rappon, and W. Berta, “Applications of artificial neural networks in
health care organizational decision-making: A scoping review,” PLOS ONE, 14(2),
e0212356, 2019. doi: 10.1371/journal.pone.0212356.
5. S. Mahadevan, P. Voigtlaender, and B. Leibe, “Iteratively Trained Interactive Seg-
mentation,” ArXiv, 2018. Available: /abs/1805.04398
6. M.Z. Zghurovskyi, N.D. Pankratova, Systems Analysis: Problems, Methodology, Ap-
plications. K.: Naukova dumka, 2005.
7. S. Walsh et al., “Decision Support Systems in Oncology,” JCO Clinical Cancer
Informatics, vol. 3, pp. 1–9, 2019. doi: 10.1200/CCI.18.00001.
8. G. Yu, Z. Chen, J. Wu, and Y. Tan, “Medical decision support system for cancer
treatment in precision medicine in developing countries,” Expert Systems with Appli-
cations, 186, 115725, 2021. doi: 10.1016/j.eswa.2021.115725.
9. T. Mukherjee, O. Pournik, S.N. Lim Choi Keung, and T.N. Arvanitis, “Clinical De-
cision Support Systems for Brain Tumour Diagnosis and Prognosis: A Systematic
Review,” Cancers 2023, 15(13), 3523. doi: 10.3390/cancers15133523.
10. B.V. Chapaliuk, Y.P. Zaychenko, “Medical image segmentation methods overview,”
System Research and Information Technologies, no. 1, pp. 72–81, 2018. doi:
10.20535/SRIT.2308-8893.2018.1.05.
11. O. Ronneberger, P. Fischer, and T. Brox, “U-Net: Convolutional Networks for Bio-
medical Image Segmentation,” ArXiv, 2015. Available: /abs/1505.04597
12. T. Lin, P. Dollár, R. Girshick, K. He, B. Hariharan, and S. Belongie, “Feature Pyra-
mid Networks for Object Detection,” ArXiv, 2016. Available: /abs/1612.03144
13. L. Chen, G. Papandreou, I. Kokkinos, K. Murphy, and A.L. Yuille, “DeepLab: Se-
mantic Image Segmentation with Deep Convolutional Nets, Atrous Convolution, and
Fully Connected CRFs,” ArXiv, 2016. Available: /abs/1606.00915
14. Z. Zhou, M.M. Siddiquee, N. Tajbakhsh, and J. Liang, “UNet++: A Nested U-Net Archi-
tecture for Medical Image Segmentation,” ArXiv, 2018. Available: /abs/1807.10165
15. H. Huang et al., “UNet 3+: A Full-Scale Connected UNet for Medical Image
Segmentation,” ArXiv, 2020. Available: /abs/2004.08790
16. A. Hatamizadeh et al., “UNETR: Transformers for 3D Medical Image Segmenta-
tion,” ArXiv, 2021. Available: /abs/2103.10504
17. Y. Zaychenko, G. Hamidov, and B. Chapaliuk, The Application of CNN and Hybrid
Networks in Medical Images Processing and Cancer Classification. Cambridge
Scholars Publishing, 2023.
18. A. Kirillov et al., “Segment Anything,” ArXiv, 2023. Available: /abs/2304.02643
19. T. Sakinis et al., “Interactive segmentation of medical images through fully convolu-
tional neural networks,” ArXiv, 2019. Available: /abs/1903.08205
Interactive decision support system for lung cancer segmentation
Системні дослідження та інформаційні технології, 2024, № 2 81
20. M. Carbonneau, V. Cheplygina, E. Granger, and G. Gagnon, “Multiple Instance
Learning: A Survey of Problem Characteristics and Applications,” Pattern Recogni-
tion, 2016. doi: 10.1016/j.patcog.2017.10.009.
21. M. Tan, Q.V. Le, “EfficientNet: Rethinking Model Scaling for Convolutional Neural
Networks,” ArXiv, 2019. Available: /abs/1905.11946
22. L. Chen, Y. Zhu, G. Papandreou, F. Schroff, and H. Adam, “Encoder-Decoder with
Atrous Separable Convolution for Semantic Image Segmentation,” ArXiv, 2018.
Available: /abs/1802.02611
23. D.P. Kingma, J. Ba, “Adam: A Method for Stochastic Optimization,” ArXiv, 2014.
Available: /abs/1412.6980
24. S. Li et al., “PyTorch Distributed: Experiences on Accelerating Data Parallel Train-
ing,” ArXiv, 2020. Available: /abs/2006.15704
25. J. Bertels et al., “Optimizing the Dice Score and Jaccard Index for Medical Image Seg-
mentation: Theory & Practice,” Medical Image Computing and Computer Assisted Inter-
vention – MICCAI 2019. doi: https://doi.org/10.1007/978-3-030-32245-8_11
26. Q. Xie, M. Luong, E. Hovy, and Q.V. Le, “Self-training with Noisy Student
improves ImageNet classification,” ArXiv, 2019. Available: /abs/1911.04252
27. M. Kimura, “Understanding Test-Time Augmentation,” in Neural Information Proc-
essing. ICONIP 2021. Lecture Notes in Computer Science, vol. 13108. Springer,
Cham, 2021. doi: https://doi.org/10.1007/978-3-030-92185-9_46
28. S.G. Armato III et al., “Data From LIDC-IDRI [Data set],” The Cancer Imaging
Archive, 2015. doi: http://10.0.31.1/K9/TCIA.2015.LO9QL9SX
29. Matthew C. Hancock, Jerry F. Magnan, “Lung nodule malignancy classification us-
ing only radiologist quantified image features as inputs to statistical learning algo-
rithms: probing the Lung Image Database Consortium dataset with two statistical
learning methods,” Journal of Medical Imaging, 3(4), 044504, 8 December 2016.
doi: https://doi.org/10.1117/1.JMI.3.4.044504
30. H. Rezatofighi, N. Tsoi, J. Gwak, A. Sadeghian, I. Reid, and S. Savarese, “General-
ized Intersection over Union: A Metric and A Loss for Bounding Box Regression,”
ArXiv, 2019. Available: /abs/1902.09630
31. Z. Wu, C. Shen, and A. van den Hengel, “Bridging category-level and instance-level
semantic image segmentation,” arXiv preprint, 2016. Available: https://
doi.org/10.48550/arXiv.1605.06885
Received 10.11.2023
INFORMATION ON THE ARTICLE
Volodymyr S. Sydorskyi, ORCID: 0000-0001-9697-7403, National Technical University
of Ukraine “Igor Sikorsky Kyiv Polytechnic Institute”, Ukraine, e-mail:
volodymyr.sydorskyi@gmail.com
ІНТЕРАКТИВНА СИСТЕМА ПІДТРИМАННЯ ПРИЙНЯТТЯ РІШЕНЬ ДЛЯ
СЕГМЕНТАЦІЇ РАКУ ЛЕГЕНІВ / В.С. Сидорський
Анотація. Досліджено клінічні інтелектуальні системи підтримання прийняття
рішень (ІСППР) для сегментації раку легень, які базуються на глибинних ней-
ронних мережах. Запропоновано нову інтерактивну ІСППР і порівняно її з по-
передніми підходами. Обговорено проблему невизначеності цілей під час
створення інтерактивних систем і запропоновано критерії для оцінювання яко-
сті та кількості зворотного зв’язку від експерта. Для автоматизації оцінювання
системи використано спеціальний алгоритм для симуляції зворотного зв’язку
експерта. Запропонована інтерактивна ІСППР перевершила попередні підходи
(як інтерактивні, так і неінтерактивні) у завданні сегментації раку легень. Цей
підхід перспективний як щодо якості, так і зручності використання експертом.
Водночас обговорено низку можливих модифікацій, які можна виконати для
покращення як критеріїв оцінювання, так і запропонованої ІСППР у майбутніх
працях.
Ключові слова: клінічні системи підтримання прийняття рішень, глибинне
навчання, відкрита система, інтерактивна сегментація.
|
| id | journaliasakpiua-article-290923 |
| institution | System research and information technologies |
| keywords_txt_mv | keywords |
| language | English |
| last_indexed | 2025-07-17T10:28:24Z |
| publishDate | 2024 |
| publisher | The National Technical University of Ukraine "Igor Sikorsky Kyiv Polytechnic Institute" |
| record_format | ojs |
| resource_txt_mv | journaliasakpiua/1e/7b0c7aa2384bb367c452b3c86feec31e.pdf |
| spelling | journaliasakpiua-article-2909232024-08-11T01:12:49Z Interactive decision support system for lung cancer segmentation Інтерактивна система підтримання прийняття рішень для сегментації раку легенів Sydorskyi, Volodymyr клінічні системи підтримання прийняття рішень глибинне навчання відкрита система інтерактивна сегментація clinical decision support systems deep learning open system interactive segmentation This paper studies Clinical Intelligent Decision Support Systems (CIDSSs) for lung cancer segmentation, which are based on deep neural nets. A new interactive CIDSS is proposed and compared with previous approaches. Additionally, the purpose uncertainty problem in building interactive systems is discussed, and criteria for measuring both quality and amount of user feedback are proposed. In order to automate system evaluation, a new algorithm was used to simulate expert feedback. The proposed interactive CIDSS outperforms previous approaches (both interactive and noninteractive) on the task of lung lesion segmentation. This approach looks promising both in terms of quality and expert user experience. At the same time, this paper discusses a bunch of possible modifications that can be done to improve both evaluation criteria and proposed CIDSS in future works. Досліджено клінічні інтелектуальні системи підтримання прийняття рішень (ІСППР) для сегментації раку легень, які базуються на глибинних нейронних мережах. Запропоновано нову інтерактивну ІСППР і порівняно її з попередніми підходами. Обговорено проблему невизначеності цілей під час створення інтерактивних систем і запропоновано критерії для оцінювання якості та кількості зворотного зв’язку від експерта. Для автоматизації оцінювання системи використано спеціальний алгоритм для симуляції зворотного зв’язку експерта. Запропонована інтерактивна ІСППР перевершила попередні підходи (як інтерактивні, так і неінтерактивні) у завданні сегментації раку легень. Цей підхід перспективний як щодо якості, так і зручності використання експертом. Водночас обговорено низку можливих модифікацій, які можна виконати для покращення як критеріїв оцінювання, так і запропонованої ІСППР у майбутніх працях. The National Technical University of Ukraine "Igor Sikorsky Kyiv Polytechnic Institute" 2024-06-28 Article Article application/pdf https://journal.iasa.kpi.ua/article/view/290923 10.20535/SRIT.2308-8893.2024.2.05 System research and information technologies; No. 2 (2024); 68-81 Системные исследования и информационные технологии; № 2 (2024); 68-81 Системні дослідження та інформаційні технології; № 2 (2024); 68-81 2308-8893 1681-6048 en https://journal.iasa.kpi.ua/article/view/290923/301102 |
| spellingShingle | клінічні системи підтримання прийняття рішень глибинне навчання відкрита система інтерактивна сегментація Sydorskyi, Volodymyr Інтерактивна система підтримання прийняття рішень для сегментації раку легенів |
| title | Інтерактивна система підтримання прийняття рішень для сегментації раку легенів |
| title_alt | Interactive decision support system for lung cancer segmentation |
| title_full | Інтерактивна система підтримання прийняття рішень для сегментації раку легенів |
| title_fullStr | Інтерактивна система підтримання прийняття рішень для сегментації раку легенів |
| title_full_unstemmed | Інтерактивна система підтримання прийняття рішень для сегментації раку легенів |
| title_short | Інтерактивна система підтримання прийняття рішень для сегментації раку легенів |
| title_sort | інтерактивна система підтримання прийняття рішень для сегментації раку легенів |
| topic | клінічні системи підтримання прийняття рішень глибинне навчання відкрита система інтерактивна сегментація |
| topic_facet | клінічні системи підтримання прийняття рішень глибинне навчання відкрита система інтерактивна сегментація clinical decision support systems deep learning open system interactive segmentation |
| url | https://journal.iasa.kpi.ua/article/view/290923 |
| work_keys_str_mv | AT sydorskyivolodymyr interactivedecisionsupportsystemforlungcancersegmentation AT sydorskyivolodymyr ínteraktivnasistemapídtrimannâprijnâttâríšenʹdlâsegmentacíírakulegenív |