Leveraging Attention Mechanisms for Interpretable Human

Embryo Image Segmentation

Wided Souid Miled

1,2

and Nozha Chakroun

LIMTIC Laboratory, Higher Institute of Computer Science, University of Tunis El-Manar, Ariana, Tunisia

National Institute of Applied Science and Technology, University of Carthage, Centre Urbain Nord, Tunisia

University of Medicine of Tunis, Lab. of Reproductive Biology and Cytogenetic, Aziza Othmana Hospital, Tunisia

Keywords:

In Vitro Fertilization, Embryo Selection, Blastocyst Segmentation, Attention Mechanisms, Deep Learning,

Explainability.

Abstract:

In-vitro Fertilization (IVF) is a widely used assisted reproductive technology where embryos are cultured un-

der controlled laboratory conditions. The selection of a high-quality blastocyst, typically reached ﬁve days

after fertilization, is crucial to the success of the IVF procedure. Therefore, evaluating embryo quality at this

stage is essential to optimize IVF outcomes. Advances in neural network architectures, particularly Convo-

lutional Neural Networks (CNNs), have enhanced decision-making in IVF. However, ensuring both accuracy

and interpretability in these models remains a challenge. This paper focuses on improving human blastocyst

segmentation by combining channel attention mechanisms with a ResNet50 model within an encoder-decoder

architecture. The method accurately identiﬁes key blastocyst components such as inner cell mass (ICM), tro-

phectoderm (TE), and zona pellucida (ZP). Our approach was validated on a publicly available human embryo

dataset, achieving Intersection over Union (IoU) scores of 83.09% for ICM, 86.87% for ZP, and 81.1% for

TE, outperforming current state-of-the-art methods. These results demonstrate the potential of deep learning

to improve both accuracy and interpretability in embryo quality assessment.

1 INTRODUCTION

In vitro fertilization (IVF) is one of the most widely

used and effective forms of assisted reproductive tech-

nology (ART) that helps couples facing fertility chal-

lenges conceive a child. Since 1978, more than 9 mil-

lion babies have been born through IVF, and approx-

imately 6% of couples experiencing infertility turn to

this procedure (Kuhnt and Passet-Wittig, 2022). The

IVF process involves several critical stages. First, ma-

ture eggs are retrieved from the ovaries and manually

combined with sperm in a controlled environment for

fertilization. The resulting fertilized egg, now called

embryo, undergoes a series of developmental phases.

Initially, the male and female pronuclei appear and

then disappear, followed by the cleavage stage, where

the single cell divides into multiple cells. Four days

after fertilization, the embryo compacts, reaching the

morula stage, and by the ﬁfth day, the embryo devel-

ops into a blastocyst.

One of the most critical steps in this complex pro-

cess is embryo selection, which aims to identify the

healthiest embryo with the highest likelihood of re-

sulting in the birth of a healthy baby. Embryo quality

is considered a key predictor of success in IVF cy-

cles. Numerous studies have demonstrated a strong

correlation between embryo morphology, implanta-

tion rates, and clinical pregnancy outcomes (Shulman

et al., 1993; Dennis et al., 2006). Consequently, se-

lecting a high-quality embryo signiﬁcantly increases

the potential for a successful pregnancy. A notewor-

thy advancement in the ﬁeld of IVF is the introduction

of time-lapse imaging incubators (TLI). This innova-

tive technology has transformed the embryo selection

process by providing a dynamic, real-time view of

embryonic development. TLI systems capture images

of each embryo at regular intervals and compile them

into a time-lapse video, offering dynamic insight into

embryonic development in vitro without disturbing

the stable culture conditions (Kovacs, 2016; Good-

man et al., 2016). Using these time-lapse videos,

embryologists grade blastocysts and identify the em-

bryos with the greatest pregnancy potential for trans-

fer to the woman’s uterus.

872

Miled, W. S. and Chakroun, N.

Leveraging Attention Mechanisms for Interpretable Human Embryo Image Segmentation.

DOI: 10.5220/0013383200003890

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 17th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2025) - Volume 2, pages 872-879

ISBN: 978-989-758-737-5; ISSN: 2184-433X

At the blastocyst stage, the embryo consists of two

main inner regions, the inner cell mass (ICM), and

the trophectoderm epithelium (TE). These regions are

surrounded by an outer membrane, the zona pellu-

cida (ZP). Both ICM and TE are considered key mor-

phological parameters for assessing embryo viability.

Therefore, evaluating the quality of these regions is

essential to determine embryonic potential. The most

commonly employed grading system for blastocyst

morphology is that of Gardner et al. (Gardner and

Schoolcraft, 1999). According to this system, three

crucial parameters are used to predict successful preg-

nancy outcomes: the degree of blastocoel cavity ex-

pansion relative to the zona pellucida, the compact-

ness of the inner cell mass, and the density of the

trophectoderm. Accurate measurement of these pa-

rameters requires segmenting human blastocyst im-

ages to delineate the three regions. Manual identiﬁca-

tion of these regions is challenging, time-consuming,

and subjective, often leading to variability between

experts. Recently, many efforts have been made to

automate embryo segmentation, utilizing both classi-

cal and deep learning-based approaches (Filho et al.,

2012; Saeedi et al., 2017; Rad et al., 2019; Muham-

mad et al., 2022). Despite its importance, embryo

segmentation still presents several challenges due to

signiﬁcant variability in the scale, shape, position, and

orientation of blastocyst components. Convolutional

Neural Networks (CNNs) have emerged as state-of-

the-art tools in medical image segmentation, offer-

ing remarkable performance in various applications.

However, conventional CNN architectures face cer-

tain limitations when applied to this task. One key is-

sue is their limited spatial awareness, particularly for

ﬂexible structures with varying shapes and positions,

as CNNs rely on shared weights in convolutional lay-

ers. This limitation can affect accurate segmenta-

tion, especially when dealing with complex, non-rigid

structures such as the inner cell mass, trophectoderm,

and zona pellucida. Moreover, the high dimensional-

ity of feature maps in CNNs can lead to redundancy,

reducing computational efﬁciency and increasing pro-

cessing time. Another signiﬁcant challenge is the lack

of interpretability in CNN decision-making. While

CNNs excel at extracting features, their black-box na-

ture makes it difﬁcult to understand why certain deci-

sions are made. This lack of explainability is a ma-

jor drawback in clinical settings, where transparent

decision-making is crucial for medical professionals

to trust and act on the results.

To address these challenges, this work focuses on

improving human embryo segmentation. As a critical

step in the IVF process, embryo segmentation enables

clinicians to assess embryo viability by delineating

key regions, including Inner Cell Mass, Zona Pellu-

cida, and Trophectoderm. Accurate segmentation of

these regions is essential for identifying embryos with

the highest implantation potential, thereby improving

IVF success rates. Building on recent advancements

in deep learning, this study proposes a novel approach

to tackle the dual challenges of segmentation accu-

racy and model interpretability.

The main contributions of this work are as follows:

• We integrate channel attention mechanisms with

a ResNet50 backbone in an encoder-decoder con-

volutional neural network. This architecture en-

hances the model’s ability to detect key hu-

man blastocyst components while providing in-

terpretable insights into the decision-making pro-

cess.

• The proposed approach is evaluated using a pub-

licly available blastocyst segmentation dataset. Its

performance is compared to state-of-the-art meth-

ods and a baseline U-Net model augmented with

the post-hoc explainability technique Grad-CAM.

• This work addresses the dual challenges of accu-

racy and interpretability in deep learning models

for embryo segmentation. By providing a frame-

work that ensures reliable segmentation while en-

hancing model transparency, it contributes to bet-

ter healthcare outcomes and fosters greater trust

in AI-driven medical solutions.

2 EXPLAINABLE AI FOR IMAGE

SEGMENTATION

Segmentation involves the automatic identiﬁcation

and delineation of speciﬁc objects or regions within

an image. The integration of explainable artiﬁcial in-

telligence (XAI) into image segmentation is gaining

attention for its potential to enhance the fairness and

reliability of AI models, particularly in sensitive do-

mains like healthcare (Saranya and Subhashini, 2023;

Gunashekar et al., 2022). In clinical applications, in-

terpretable segmentation models are essential to build

trust, improve communication with patients, and sup-

port iterative reﬁnements to achieve more precise and

accurate results. This section explores XAI methods

designed to clarify the decision-making processes of

segmentation models, a critical step toward enhancing

their applicability in clinical contexts and advancing

research efforts.

Leveraging Attention Mechanisms for Interpretable Human Embryo Image Segmentation

873

2.1 Ad-Hoc Explainability Methods

Ad-hoc explainability methods are external tech-

niques applied to pre-existing AI models to provide

insight into their decision-making processes. These

methods are particularly valuable for complex mod-

els, such as deep learning architectures, which in-

herently lack transparency. Commonly used ad hoc

techniques for image segmentation include SHapley

Additive exPlanations (SHAP), Local Interpretable

Model-agnostic Explanations (LIME), and Gradient-

weighted Class Activation Mapping (Grad-CAM).

SHAP assigns contribution scores to input features,

quantifying their inﬂuence on the model’s predictions.

By leveraging game theory, SHAP ensures fair attri-

bution of prediction credit, providing a detailed un-

derstanding of the importance of the feature. LIME

generates interpretable surrogate models to approxi-

mate the behavior of a complex model around a spe-

ciﬁc prediction. This helps explain individual predic-

tions, especially in ambiguous or error-prone cases.

The Grad-CAM approach (Selvaraju et al., 2017;

Vinogradova et al., 2020) utilizes gradients from the

ﬁnal convolutional layer to highlight image regions

that most inﬂuence the model’s predictions, provid-

ing intuitive visual explanations. Its enhanced vari-

ant, Grad-CAM++, incorporates information from all

convolutional layers, reﬁning the localization of dis-

criminative regions and delivering more precise ex-

planations.

2.2 Self-Explainable Methods

Unlike ad-hoc methods, self-explainable models are

designed with interpretability as an integral part of

their architecture. These models inherently inte-

grate explainability into their structure, making their

decision-making processes clear. One prominent ex-

ample is the ProtoSeg method, proposed by Sacha

et al. (Sacha et al., 2023), which uses prototype-

based learning to assign pixels to classes by compar-

ing image regions to prototypes, which are character-

istic examples of each class. This approach naturally

enhances interpretability, as segmentation decisions

are explicitly linked to these prototypes. Attention-

based models leverage attention mechanisms to focus

on relevant regions within an image during segmen-

tation (Gu et al., 2020; Zhao et al., 2021). By as-

signing importance scores to features, they generate

interpretable attention maps, simultaneously improv-

ing segmentation accuracy and transparency.

The choice between ad hoc and self-explainable

methods depends on the speciﬁc requirements of the

applications. Ad-hoc methods are suitable for provid-

ing detailed and ﬂexible analysis, making them ideal

for in-depth evaluations of existing models. How-

ever, they often require technical expertise and com-

putational resources. In contrast, self-explainable

methods prioritize intuitive interpretation and com-

putational efﬁciency, making them well-suited for

resource-constrained environments or users with lim-

ited technical expertise.

3 METHODOLOGY

3.1 Model Architecture

3.1.1 Overview

The proposed Residual Attention Network (ResA-

Net), inspired by the method introduced in (Gu et al.,

2020), integrates comprehensive attention mecha-

nisms for efﬁcient and interpretable image segmen-

tation. While retaining the general structure of the

original model, we enhanced its backbone by replac-

ing it with ResNet50 to leverage its superior feature

extraction capabilities. Specialized attention modules

are incorporated to guide segmentation across spatial,

channel, and scale dimensions simultaneously. The

architecture comprises four spatial attention modules

(SA1-SA4), four channel attention modules (CA1-

CA4), and a scale attention module (LA), each con-

tributing to precise and context-aware segmentation.

ResA-Net adopts an encoder-decoder architec-

ture. The encoder processes input images through

convolutional layers and max-pooling operations, re-

ducing spatial dimensions while capturing high-level

and multi-scale features. These features are passed

to a central layer, bridging the encoder and decoder,

before being fed into the decoder to reconstruct the

segmentation output. The decoder employs bilinear

interpolation and concatenation to upsample feature

maps to the original resolution while effectively com-

bining information from multiple scales. Finally, a

1×1 convolutional layer reduces feature map dimen-

sions, and a class-speciﬁc convolutional layer gener-

ates probability maps for each segmentation class. A

softmax layer converts these probabilities into the ﬁ-

nal segmentation output by assigning each pixel to the

class with the highest likelihood. An overview of the

proposed method is presented in Figure 1.

While achieving accurate segmentation is essen-

tial, our focus extends beyond the segmentation task

to understanding the model’s decision-making pro-

cess. To this end, we emphasize the visualization of

attention maps generated by ResA-Net, which high-

light inﬂuential regions in the input image. These

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

874

Figure 1: Overview of the proposed ResA-Net model.

visualizations offer valuable insights into how the

model makes its decisions, enhancing transparency

and interpretability by clearly showcasing the regions

and features that guide its segmentation outcomes.

3.1.2 Attention Mechanisms in ResA-Net

ResA-Net employs three attention mechanisms Spa-

tial Attention Modules (SA), Channel Attention Mod-

ules (CA), and a Scale Attention Module (LA).

Spatial Attention: Inspired by the non-local net-

work and Attention Gates (AG), ResA-Net uses four

spatial attention blocks (SA1-SA4) to learn atten-

tion maps across different resolution levels. These

maps emphasize relevant regions of the input image

while suppressing noise, enhancing segmentation ac-

curacy. At the lowest resolution level (SA1), a non-

local block captures global context, while at higher

levels (SA2-SA4), attention gates focus on key areas

within the image. Low-level spatial attention features

from the encoder are concatenated with high-level de-

coder features to enhance segmentation through com-

plementary information.

Channel Attention: Channel attention modules au-

tomatically identify and amplify relevant feature

channels while suppressing irrelevant ones. This

mechanism ensures that the model focuses on seman-

tically meaningful features crucial for accurate seg-

mentation.

Scale Attention: To handle variations in object

scale, the scale attention module assigns adaptive

weights to features at different scales. This allows

the network to prioritize features most relevant to the

speciﬁc image, improving segmentation performance.

By combining these mechanisms, ResA-Net not only

achieves high segmentation accuracy but also pro-

vides interpretable outputs, offering insights into the

decision-making process and enabling its application

in sensitive ﬁelds such as healthcare.

3.2 Dataset

The dataset employed in this paper for segmenting

human embryo images is the unique publicly avail-

able dataset introduced by Saaedi et al. (Saeedi et al.,

2017). It comprises 235 blastocyst images from pa-

tients treated at the Paciﬁc Center of Reproductive

Medicine in Canada between 2012 and 2016. Each

image was manually annotated by expert embryolo-

gists and provided with ground truth binary masks

relating to ICM, TE, and ZP regions. Using this

dataset, we employed the proposed model architec-

ture to train three distinct conﬁgurations, each fo-

cused on segmenting a speciﬁc region of interest. This

work adopts a single-class segmentation approach to

individually identify the ZP, ICM, and TE regions

within blastocyst images. By generating explicit seg-

mentation labels, the model facilitates both accurate

segmentation and enhanced interpretability analysis.

4 EXPERIMENTAL RESULTS

4.1 Implementation Details

A consistent preprocessing pipeline was applied to all

three one-class segmentation conﬁgurations to ensure

uniformity and efﬁciency. All images and their corre-

Leveraging Attention Mechanisms for Interpretable Human Embryo Image Segmentation

875

sponding ground truth masks were resized to a ﬁxed

resolution of 224× 300, to ensure consistent input di-

mensions for the model. Both the images and masks

were then converted to .npy format to facilitate efﬁ-

cient storage and handling during training.

To manage the dataset, we implemented a custom

dataset class using PyTorch’s Dataset module. This

class was speciﬁcally designed to pair image sam-

ples with their corresponding labels, with each dataset

entry represented as tensors loaded from the prepro-

cessed .npy ﬁles. The dataset was divided into train-

ing (85%), and testing (15%) subsets, ensuring a bal-

anced evaluation of the model’s performance. Data

loading was performed using PyTorch’s DataLoader,

enabling efﬁcient batching and streamlined access to

the dataset. A batch size of 16 images was selected to

balance memory constraints and training efﬁciency.

The Adam optimizer was employed with a learning

rate of 10

−4

and a weight decay of 10

−8

, to opti-

mize the model parameters. The Dice loss function

was utilized to evaluate segmentation performance,

as it effectively handles the class imbalance inherent

in medical imaging datasets. Each conﬁguration was

trained for 100 epochs to ensure convergence and op-

timal performance. Training was conducted locally

on a desktop PC equipped with an NVIDIA GeForce

GTX 1650 graphics card, providing sufﬁcient compu-

tational power for the segmentation task.

4.2 Evaluation

For the performance evaluation of the proposed

model, we used the Intersection over Union (IoU),

also known as the Jaccard Index, and the F1 score.

The Jaccard Index is the ratio of the intersection to

the union of the segmentation result and the ground

truth. The F1 score is deﬁned as twice the area

of overlap between the segmentation result and the

ground truth, divided by the sum of the areas of

both. Table 1 presents a quantitative comparison

of the segmentation performance between the pro-

posed method, the baseline U-Net model and recent

state-of-the art methods. U-Net (Ronneberger et al.,

2015) has demonstrated its effectivenes in segmenting

anatomical structures across various imaging modali-

ties, including MRI, CT scans, and microscopy. Blas-

Net (Rad et al., 2019) enhances segmentation by in-

corporating multi-scale global contextual information

through a cascaded atrous pyramid pooling module

and reproducing feature resolution using dense pro-

gressive sub-pixel upsampling. MASS-Net (Muham-

mad et al., 2022) leverages depth-wise concatenation

to combine spatial information at multiple scales, en-

abling the simultaneous detection of blastocyst com-

ponents. ECS-Net (Mushtaq et al., 2022) employs

dual streams: base convolutional and depth-wise sep-

arable convolutional blocks, densely concatenated to

enrich features for improved segmentation. FSBS-

Net (Ishaq et al., 2023) introduces feature supple-

mentation across scales and employs ascending chan-

nel convolutional blocks (ACCB) to reﬁne blastocyst

segmentation with minimal computational overhead,

leveraging skip connections to integrate features from

shallow and deep layers.

Based on the Intersection over Union (IoU) met-

ric, the ResA-Net model consistently outperforms

state-of-the-art models in segmenting the ZP and TE

regions. While its performance on the ICM region

is slightly below that of some state-of-the-art mod-

els, ResA-Net achieves outstanding results in terms of

the mean IoU across all three regions. This demon-

strates the model’s overall effectiveness and robust-

ness in segmentation tasks.

4.3 Models Interpretability

4.3.1 Visualization with Grad-CAM for U-Net

To better understand the areas of interest that guide

the U-Net model’s segmentation decisions, we used

Grad-CAM to generate heatmaps for each segmen-

tation class. These heatmaps were then combined

into a single heatmap that highlights the regions the

model focuses on during the segmentation process.

Grad-CAM was applied to the last convolutional layer

of the U-Net model, providing insight into how the

model attends to different regions of the image. For

the ZP region, the heatmap shows clear focus on spe-

ciﬁc areas within the embryo’s perimeter, which is

the most straightforward to segment. In contrast, for

the ICM region, the model appears to focus on a sin-

gle, central entity within the embryo. This suggests

that the model identiﬁes a distinct, central structure to

guide segmentation. The Trophectoderm (TE) region,

which has the lowest segmentation metric scores, is

the most challenging to identify. The Grad-CAM

heatmap reveals that the model attempts to identify

multiple entities within the inner perimeter of the em-

bryo, which later converge to form the TE region. Fig-

ure 2 presents examples of the combined Grad-CAM

outputs applied to the ﬁnal layers of the three U-Net

conﬁgurations, illustrating the model’s focus for each

segmentation class.

4.3.2 Visualization of ResA-Net Attention Maps

Figure 3 presents the attention maps generated by the

ResA-Net model for each segmentation conﬁguration

(ICM, ZP, TE), derived from the last image in the ﬁ-

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

876

Table 1: Comparative analysis of proposed method with state-of-the-art models in terms of IoU and F1 score.

ZP TE ICM Mean

Model IoU F1 IoU F1 IoU F1 IoU

U-Net Baseline 79.10 88.33 71.66 83.49 76.33 86.57 75.69

BlastNet (Rad et al., 2019) 81.15 - 76.52 - 81.07 - 79.58

MASS-Net (Muhammad et al., 2022) 84.69 - 79.08 - 85.88 - 83.22

ECS-Net (Mushtaq et al., 2022) 85.34 - 78.43 - 85.26 - 83.01

FSBS-Net (Ishaq et al., 2023) 85.80 92.29 80.17 88.90 85.55 92.0 83.84

ResA-Net 86.87 92.72 81.10 88.88 83.09 90.09 83.68

Figure 2: GradCAM heatmaps for U-Net model.

nal batch of the training dataset. These maps offer

valuable insights into the model’s segmentation pro-

cess for each class. In the case of the Zona Pellucida

one-class segmentation conﬁguration, the attention

maps reveal a precise and systematic segmentation

process. The ﬁrst map clearly delineates the ZP from

its surrounding background, while subsequent maps

progressively reﬁne the segmentation by emphasiz-

ing key feature channels, particularly highlighting the

boundary between the ZP and TE regions. This pro-

gression helps accurately deﬁne the contours separat-

ing these regions, showcasing the model’s ability to

focus on the critical features needed for precise seg-

mentation.

For the Inner Cell Mass one-class conﬁguration,

the attention maps demonstrate how the model iden-

tiﬁes the relevant regions within the blastocyst. The

initial map shows the model’s broad understanding of

the entire blastocyst, detecting it as a whole. The sec-

ond map narrows in on the ICM, accurately localizing

this important region. This sequential reﬁnement in-

dicates the model’s capacity to differentiate between

the various components of the blastocyst, enhancing

segmentation precision. However, a slight overlap is

observed between the ICM and TE regions, where

some TE pixels are also highlighted. While this mi-

nor ambiguity does not signiﬁcantly affect the overall

performance, it underscores the challenges of distin-

guishing adjacent regions.

In contrast, the attention maps for the Trophec-

toderm one-class conﬁguration reveal challenges in

segmentation. Unlike the ZP and ICM models, the

TE model struggles to isolate the TE region. The ini-

tial attention maps fail to clearly differentiate the TE

from the background or the blastocyst structure. De-

spite relatively high segmentation metrics for the TE

region, these scores are lower compared to those for

ZP and ICM, suggesting that this particular input im-

age may not fully represent the model’s performance

across the entire dataset.

5 CONCLUSION

In IVF procedures, accurate segmentation is essen-

tial for assessing embryo viability. By dividing the

embryo into distinct regions, such as the Inner Cell

Mass, Zona Pellucida, and Trophectoderm, clinicians

can better identify embryos suitable for implanta-

tion, ultimately improving the chances of a success-

ful pregnancy. This paper introduces the development

of three one-class segmentation models for embryo

image analysis, incorporating attention mechanisms

with a ResNet50 backbone in an encoder-decoder

architecture. Experimental results demonstrate that

the integration of attention mechanisms enhances the

model’s ability to focus on relevant regions, leading

to improved segmentation performance.

To further enhance segmentation results, future

work could focus on reﬁning the current models by in-

corporating additional contextual information, which

would improve the differentiation of closely related

regions. Additionally, exploring more advanced at-

tention mechanisms and hybrid architectures, com-

bining CNNs with transformer models, could cap-

ture ﬁne-grained details more effectively, ultimately

boosting performance.

Leveraging Attention Mechanisms for Interpretable Human Embryo Image Segmentation

877

Figure 3: Attention Heatmaps Generated by the ResA-Net Model.

REFERENCES

Dennis, S. J., Thomas, M. A., Williams, D. B., and Robins,

J. C. (2006). Embryo morphology score on day 3 is

predictive of implantation and live birth rates. Journal

of assisted reproduction and genetics, 23(4).

Filho, E., Noble, J., Poli, M., Grifﬁths, T., Emerson, G., and

Wells, D. (2012). A method for semi-automatic grad-

ing of human blastocyst microscope images. Human

reproduction (Oxford, England), 27:2641–2648.

Gardner, D. K. and Schoolcraft, W. B. (1999). Culture and

transfer of human blastocysts. Current Opinion in Ob-

stetrics and Gynaecology, 11(3):307–311.

Goodman, L. R., Goldberg, J., Falcone, T., Austin, C., and

Desai, N. (2016). Does the addition of time-lapse

morphokinetics in the selection of embryos for trans-

fer improve pregnancy rates? a randomized controlled

trial. Journal of Fertility and Sterility, 105(2).

Gu, R., Wang, G., Song, T., Huang, R., Aertsen, M., De-

prest, J., Ourselin, S., Vercauteren, T., and Zhang, S.

(2020). Ca-net: Comprehensive attention convolu-

tional neural networks for explainable medical image

segmentation. IEEE transactions on medical imaging,

40(2):699–711.

Gunashekar, D. D., Bielak, L., H

agele, L., Oerther, B.,

Benndorf, M., Grosu, A., Brox, T., Zamboglou, C.,

and Bock, M. (2022). Explainable ai for cnn-based

prostate tumor segmentation in multi parametric mri

correlated to whole mount histopathology. Radiation

Oncology, 17(1):65.

Ishaq, M., Raza, S., Rehar, H., Abadeen, S., Hussain, D.,

Naqvi, R., and Lee, S. W. (2023). Assisting the human

embryo viability assessment by deep learning for in

vitro fertilization. Mathematics, 11(9):2023.

Kovacs, P. (2016). Time-lapse embryoscopy: Do we have

an efﬁcacious algorithm for embryo selection? Jour-

nal of Reproductive Biotechnology and Fertility, 5.

Kuhnt, A. K. and Passet-Wittig, J. (2022). Families formed

through assisted reproductive technology: Causes, ex-

periences, and consequences in an international con-

text. In Reprod Biomed Soc Online, number 14, pages

289–296.

Muhammad, A., Adnan, H., Woon, C. S., Hwan, K. Y.,

and Ryoung, P. K. (2022). Human blastocyst com-

ponents detection using multiscale aggregation se-

mantic segmentation network for embryonic analysis.

Biomedicines, 10(7):1717.

Mushtaq, A., Mumtaz, M., Raza, A., Salem, N., and Yasir,

M. N. (2022). Artiﬁcial intelligence-based detection

of human embryo components for assisted reproduc-

tion by in vitro fertilization. Sensors, 22(19):7418.

Rad, R. M., Saeedi, P., Au, J., and Havelock, J. (2019).

Blast-net: Semantic segmentation of human blasto-

cyst components via cascaded atrous pyramid and

dense progressive upsampling. In IEEE International

Conference on Image Processing (ICIP), pages 1865–

1869.

Ronneberger, O., Fischer, P., and Brox, T. (2015). U-

net: Convolutional networks for biomedical image

segmentation. In Medical image computing and

computer-assisted intervention–MICCAI 2015: 18th

international conference, Munich, Germany, October

5-9, 2015, proceedings, part III 18, pages 234–241.

Springer.

Sacha, M., Rymarczyk, D., Struski, L., Tabor, J., and

Zieli

nski, B. (2023). Protoseg: Interpretable semantic

segmentation with prototypical parts. In Proceedings

of the IEEE/CVF Winter Conference on Applications

of Computer Vision, pages 1481–1492.

Saeedi, P., Yee, D., Au, J., and Havelock, J. (2017). Auto-

matic identiﬁcation of human blastocyst components

via texture. IEEE Transactions on Biomedical Engi-

neering, 64(12):2968–2978.

Saranya, A. and Subhashini, R. (2023). A systematic re-

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

878

view of explainable artiﬁcial intelligence models and

applications: Recent developments and future trends.

Decision analytics journal, page 100230.

Selvaraju, R., Cogswell, M., Das, A., Vedantam, R., Parikh,

D., and Batra, D. (2017). Grad-cam: Visual explana-

tions from deep networks via gradient-based localiza-

tion. In Proceedings of the IEEE international confer-

ence on computer vision, pages 618–626.

Shulman, A., Ben-Nun, I., Y. Ghetler, H. K., Shilon, M.,

and Beyth, Y. (1993). Relationship between embryo

morphology and implantation rate after in vitro fertil-

ization treatment in conception cycles. Fertility and

sterility, 60(1).

Vinogradova, K., Dibrov, A., and Myers, G. (2020).

Towards interpretable semantic segmentation via

gradient-weighted class activation mapping (student

abstract). Proceedings of the AAAI Conference on Ar-

tiﬁcial Intelligence, 34(10):13943–13944.

Zhao, Q., Liu, J., Li, Y., and Zhang, H. (2021). Seman-

tic segmentation with attention mechanism for remote

sensing images. IEEE Transactions on Geoscience

and Remote Sensing, PP:1–13.

Leveraging Attention Mechanisms for Interpretable Human Embryo Image Segmentation

879