Bias Assessment in Medical Imaging Analysis:

A Case Study on Retinal OCT Image Classiﬁcation

Gabriel Oliveira, Lucas David, Rafael Padilha, Ana Paula da Silva,

Francine de Paula, Lucas Infante, Lucio Jorge, Patricia Xavier and Zanoni Dias

Institute of Computing, University of Campinas, Campinas, SP, Brazil

Keywords:

Deep Learning, Dataset Bias, Model Interpretability, Medical image diagnosis, Retinal OCT Analysis.

Abstract:

Deep learning classiﬁers can achieve high accuracy in many medical imaging analysis problems. However,

when evaluating images from outside the training distribution — e.g., from new patients or generated by

different medical equipment — their performance is often hindered, highlighting that they might have learned

speciﬁc characteristics and biases of the training set and can not generalize to real-world scenarios. In this

work, we discuss how Transfer Learning, the standard training technique employed in most visual medical

tasks in the literature, coupled with small and poorly collected datasets, can induce the model to capture such

biases and data collection artifacts. We use the classiﬁcation of eye diseases from retinal OCT images as the

backdrop for our discussion, evaluating several well-established convolutional neural network architectures

for this problem. Our experiments showed that models can achieve high accuracy in this problem, yet when

we interpret their decisions and learned features, they often pay attention to regions of the images unrelated to

diseases.

1 INTRODUCTION

With data-driven approaches achieving promising

results in many inference tasks, numerous deep

learning-based methods were proposed in the past

decade for the medical domain (Litjens et al., 2017).

Many works focus on obtaining highly accurate mod-

els for tasks such as diagnostics, medical imaging

analysis, referral assessment, drug discovery — all of

which could signiﬁcantly improve healthcare in our

society. Although important, accuracy alone is not

enough for an automatic approach to be applied in a

real scenario in which its decision might affect peo-

ple’s lives. In such cases, it is essential to guarantee

the transparency and interpretability of the model, as-

sessing the factors that inﬂuence each automatic an-

swer. Doing so improves the trustworthiness of the

algorithm, which is pivotal for its acceptance in prac-

tical scenarios.

When transparency is overlooked, black-box

models might output correct answers for the wrong

reasons. A recent example of this occurred during

the COVID-19 outbreak. Moved by the urgency of

the pandemic, a large body of works proposed image-

based diagnostic methods that reported high accuracy

in diverse scenarios (Shi et al., 2020). However, in

a recent analysis (Roberts et al., 2021), the authors

evaluated a pool of 415 works proposing machine

learning-based approaches for COVID-19 diagnostic

through chest X-ray and computational tomography

scans. According to their assessment, none of the

works could be used in clinical practice, primarily due

to biases and inference ﬂaws learned by the model

during training. Such undesired effects may originate

from methodological errors during dataset collection

and sanitization that are mistakenly leveraged during

model optimization. For example, if images from a

patient present a frequent acquisition artifact that dis-

tinguishes them from other patients’, the model might

learn to identify that artifact to classify a disease in-

stead of learning the correct features for a diagnosis.

Deep learning models often consist of complex

and parameter-heavy neural networks able to directly

learn the most discriminative characteristics for the

target problem from available data. To properly learn

them, most models require vast amounts of annotated

data, which are not always available, especially in the

medical domain. To overcome such limitations, re-

searchers employ Transfer Learning, pre-training the

model on a different domain with plenty of data and

further ﬁne-tuning the acquired knowledge to the tar-

get task. However, when trained with datasets whose

collection and sanitization were not rigorously per-

formed to mitigate possible biases (e.g., lack of pa-

tient and sensor representativity, data leakage during

the organization of training and validation splits), the

574

Oliveira, G., David, L., Padilha, R., Paula da Silva, A., de Paula, F., Infante, L., Jorge, L., Xavier, P. and Dias, Z.

Bias Assessment in Medical Imaging Analysis: A Case Study on Retinal OCT Image Classiﬁcation.

DOI: 10.5220/0010867400003116

In Proceedings of the 14th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2022) - Volume 3, pages 574-580

ISBN: 978-989-758-547-0; ISSN: 2184-433X

optimization may lead the model to learn whichever

features better solve the task, regardless of their qual-

ity. Consequently, this behavior might reﬂect in an

artiﬁcially high accuracy that does not generalize to

real application scenarios. Once trained, it becomes

hard to identify and ﬁx such issues due to their black-

box nature and the insufﬁcient reproducibility details

presented by most works.

In this work, we extend the discussion about

dataset biases and how modern data-driven tech-

niques are prone to capture them instead of focusing

on the task at hand. We examine the training proce-

dure frequently resorted by deep learning-based ap-

proaches and discuss the characteristics of medical

datasets employed in their training that might lead to

low generalization at inference time. To better exem-

plify these issues and what type of artifacts and bi-

ases are captured in a real scenario, we analyze convo-

lutional neural networks (CNN) to diagnose eye dis-

eases on retinal optical coherence tomography (OCT)

scans. Finally, we discuss interpretability techniques

that can aid us in identifying and mitigating biases

captured during training.

The remainder of the work is organized as follows.

In Section 2, we discuss how researchers leverage

datasets from other domains with the Transfer Learn-

ing technique, as well as the issues and biases that

originated during dataset collection and sanitization

that might affect generalization. Whereas, in Sec-

tion 3, we present the case analysis of retinal OCT

classiﬁcation, evaluating several deep learning mod-

els under an experimental scenario and interpreting

their decisions. Finally, in Section 4, we discuss our

ﬁnal thoughts, highlighting the importance of explain-

ability approaches for bias mitigation.

2 RELATED CONCEPTS

Convolutional Neural Networks have achieved

promising results in different visual domains, includ-

ing medical problems (Deepak and Ameer, 2019;

Oliveira et al., 2020). Due to their data requirement

and the lack of annotated data in medical tasks, most

methods in the literature employ Transfer Learning as

the standard approach to deal with smaller datasets.

The technique originated from educational psychol-

ogy, which states that experiences in one domain

can be generalized to another (Zhuang et al., 2021).

This approach is widely used by the deep learning

community as it alleviates the data requirements for

complex models whose training from scratch would

be unfeasible (Tan et al., 2018).

The rationale behind Transfer Learning is that,

once optimized, the initial and intermediate layers of

CNNs tend to capture low-level information in im-

ages, such as edges, corners, and color blobs. These

superﬁcial representations are often shared between

tasks of distinct visual domains (Yosinski et al., 2014;

Hussain et al., 2018) — from object classiﬁcation

to medical imaging analysis — allowing them to be

transferred for new tasks. On the other hand, deeper

layers build on top of the low-level concepts, learning

domain-speciﬁc characteristics specialized in the task

at hand that cannot be applied to other problems.

The pre-trained models are optimized in large an-

notated datasets before having their features trans-

ferred. Most works pre-train their approaches on

ImageNet (Deng et al., 2009), an object classiﬁca-

tion dataset with 14 million images and 1000 classes.

The adaptation and adjustment to the new domain are

made by ﬁne-tuning the weights of the CNN with the

target dataset, using the pre-trained weights as the

starting point for the optimization. This is usually

done with a lower learning rate, as the objective is

solely to tweak (and not to re-train) the network to

generalize for the new domain.

Even though the training procedure plays a crucial

part in the performance of the ﬁnal model, it is only

one of its critical components. The quality of data

has a signiﬁcant impact on the generalization of deep

learning models. Due to the sensitive nature of med-

ical tasks, biases and ﬂaws introduced in the dataset

during its collection and sanitization could have se-

vere repercussions during inference in real-world sce-

narios. As discussed by previous works (Roberts

et al., 2021), dataset bias has many possible origins

that should be considered, but most of them tend to

fall under the lack of data representativity, ﬂaws in the

collection process, and data leakage or contamination

during training and evaluation.

Medical images are captured with expensive

equipment (e.g., computed tomography, ultrasound,

and optical coherence tomography) that can differ in

characteristics and parameters from one machine to

another even when considering similar models (Wu

et al., 2018). Additionally, images are captured from

several patients, usually following a protocol oriented

by a technician. Issues with the equipment (e.g., sen-

sor noises from a particular machine), in the collec-

tion procedure (e.g., the person moving during the

exam), and patient proﬁle (e.g., imagery captured

from a particular age and gender), if not accounted,

can all introduce artifacts that might be exploited by

the model afterward.

Besides those, several issues can be introduced

when organizing the dataset into training and testing

Bias Assessment in Medical Imaging Analysis: A Case Study on Retinal OCT Image Classiﬁcation

575

splits, as well as posteriorly during the model evalu-

ation. Data leakage is a methodological mistake that

can have subtle consequences. For example, having

shared patients between splits might induce the model

to recognize the patient’s identity (through body char-

acteristics or how that person posed for capture) in-

stead of focusing on the disease features. A similar

issue can happen when the data from one class of

our task is all captured by the same equipment that

was not used for patients of other classes, allowing

the model to correlate the sensor noise of the machine

with the diagnostic.

3 CASE ANALYSIS:

CLASSIFICATION OF

RETINAL OCT IMAGES

Vision impairment is a growing medical concern in

our society,

in which early diagnosis plays an im-

portant role in prevention and treatment assessment.

Approximately 30 million OCT scans are performed

each year worldwide(Swanson and Fujimoto, 2017),

requiring extensive human supervision to ﬁlter and

analyze potential patients. Considering this, auto-

matic medical imaging analysis methods are becom-

ing important tools to process the high volume of

scans in a timely and effective manner.

In this section, we evaluate several CNN architec-

tures over a scenario of classiﬁcation of eye diseases

through the analysis of retinal OCT scans. Firstly, we

describe the employed dataset and its properties. We

then discuss the methodology and the experimental

evaluation setup. Finally, we present the results and

interpret them using explainability techniques.

3.1 Dataset

The dataset used in this work was collected from pa-

tients from several hospitals and ophthalmology in-

stitutes in the USA and China between 2013 and

2017 (Kermany et al., 2018b; Kermany et al., 2018a).

Each OCT scan belongs to one of four classes:

Choroidal Neovascularization (CNV), Diabetic Mac-

ular Edema (DME), Drusen, and Normal. Images as-

signed as Normal indicate that they belong to healthy

patients without any sign of diseases, such as ﬂuids

or edemas. Figure 1 presents examples of each avail-

able class. Several OCT images in the dataset present

some type of noise, such as in- and out-of-plane ro-

tations and image shearing, as presented in Figure 2.

https://www.who.int/en/news-room/fact-sheets/detail/

blindness-and-visual-impairment

(a) (b)

Figure 1: Examples of OCT scans from (a) Choroidal

Neovascularization (CNV), (b) Diabetic Macular Edema

(DME), (c) Drusen, and (d) Normal classes.

Figure 2: Examples of noisy OCT images, with different

degrees of rotation, crop and shearing.

Table 1: OCT images distribution on training, validation,

and testing sets. Class distribution is unbalanced on the

training and validation sets, but balanced on the testing set.

Class Train Validation Test

CNV 19,115 3,245 242

DME 5,484 1,414 242

Drusen 3,269 590 242

Normal 18,214 4,466 242

Total 46,082 9,715 968

Besides that, images vary in resolution, ranging from

512 × 512 up to 1536 × 496.

The dataset has already been organized into train-

ing, validation and testing splits. However, when ana-

lyzing the split composition, we noticed that some pa-

tients had images in multiple sets. By sharing patients

between the sets, the models might be encouraged

to learn patient-speciﬁc characteristics instead of dis-

criminative features of the diseases. Even though this

might lead to a higher classiﬁcation accuracy on the

testing set, it is an undesired effect as it hinders their

ICAART 2022 - 14th International Conference on Agents and Artiﬁcial Intelligence

576

CNN

CNV

DME

Drusen

Normal

Classification

Figure 3: Pipeline of our method.

generalization to scans from new patients. To avoid

data leakage and not harm the learning capability of

the models, we redeﬁned the training and validation,

removing the images from patients that are also pre-

sented on the testing set. We gathered all the images

from the training and validation sets and divided them

into 80% for training and 20% for validation, ensur-

ing that the same patient was present in only one of

the sets. Table 1 shows the number of OTC scans on

training, validation, and testing sets with the new con-

ﬁguration.

3.2 Methodology

In this subsection, we present our approach for the

problem of classiﬁcation of retina-related diseases.

The pipeline of our method is presented in Figure 3.

Firstly, we employ well-established CNN archi-

tectures, pre-trained over the ImageNet dataset (Deng

et al., 2009), to perform feature extraction. The

output features are summarized with a Global Aver-

age Pooling layer (GAP) and used to train a dense

softmax classiﬁer. We experiment with the fol-

lowing architectures: ResNet-50 (He et al., 2016),

MobileNet (Howard et al., 2017), VGG-16 and

VGG-19 (Simonyan and Zisserman, 2014), Efﬁcient-

NetB0 (Tan and Le, 2019), and InceptionV3 (Szegedy

et al., 2016). For each feature extracting network, the

softmax classiﬁer is trained for ten epochs using Nes-

terov Momentum SGD optimizer for the categorical

cross-entropy loss function. We consider a learning

rate of 0.001 and a momentum factor of 0.9. Also, we

use the early stopping technique to halt the training

after four epochs without decreasing the loss function

on the validation set.

In a second phase, we select the topmost scoring

architecture from the previous step and ﬁne-tune it

over the Retinal OCT dataset. Unlike the previous

step, in which the networks’ weights were ﬁxed, we

allow the CNN to update its parameters to match the

training OCT images to their respective labels. The

same training procedure, as well as hyperparameters,

are employed in this phase.

In both experiments, we augment the training data

using these three operations: random horizontal ﬂip,

random zoom in the range of [0.8, 1.2], and random

(a) (b)

Figure 4: Examples of data augmentation: (a) original im-

age, (b) horizontal ﬂip, (c) zoom, and (d) shear.

shear in the range of [0.8, 1.2]. Figure 4 shows these

three operations of data augmentation.

3.3 Experimental Evaluation

In this subsection, we present the experimental eval-

uation of different CNN architectures considering

transfer learning and ﬁne-tuning techniques. Firstly,

we report the feature extraction results using CNNs

trained over the ImageNet dataset. Then we report

and discuss the results of the ﬁne-tuning experiment

and, ﬁnally, we evaluate the highest performing net-

work against the test set. Finally, we consider the in-

terpretability technique Grad-CAM (Selvaraju et al.,

2017) to evaluate the features being utilized by the

model’s decision process.

3.3.1 Transfer Learning with ImageNet Weights

The result of each architecture is presented in Ta-

ble 2. ResNet50 reached the highest balanced accu-

racy on the validation set between the selected archi-

tectures, with a score of 79.75%. The remaining net-

works achieved similar results, with InceptionV3 out-

performed by all of them.

3.3.2 Fine-tuned Network

With our results considering the convolutional neural

network architectures on the validation set with trans-

Bias Assessment in Medical Imaging Analysis: A Case Study on Retinal OCT Image Classiﬁcation

577

Table 2: Results of transfer learning and ﬁne-tuning on the validation set. ResNet50 (He et al., 2016) with transfer learning

achieved the highest balanced accuracy on the validation set. With that, we ﬁne-tuned this architecture, outperforming the

previous results with transfer learning.

Network Balanced Accuracy (%)

Transfer Learning

ResNet50 (He et al., 2016) 79.75

MobileNet (Howard et al., 2017) 79.68

VGG-16 (Simonyan and Zisserman, 2014) 79.21

VGG-19 (Simonyan and Zisserman, 2014) 78.96

EfﬁcientNetB0 (Tan and Le, 2019) 78.32

InceptionV3 (Szegedy et al., 2016) 75.98

Fine-tuning

ResNet50 (He et al., 2016) 89.95

Figure 5: Progress of the loss function for the training and

validation sets throughout the optimization of the ﬁne-tuned

ResNet50.

fer learning technique, we performed further investi-

gations with the best CNN.

The best result was reached by ResNet50, which

achieved 79.75% of balanced accuracy on the valida-

tion set. With that, we employed the ﬁne-tuning tech-

nique, i.e., we retrained ResNet50, initialized with

ImageNet weights, but allowing them to be freely up-

dated during optimization.

Added to the ﬁne-tuning process, we analyzed the

loss function during the training step, as shown in

Figure 5. The loss curve of training and validation

phases indicates that the training converged quickly

to a global minimum in the training set. However, the

validation loss stayed close to the same point, stop-

ping the training process due to the early-stopping

technique. Considering the curve, we can highlight

the impact of the hyperparameters chosen for the

SGD optimizer, and further investigations can be done

with different values for the hyperparameters.

As a result, the ﬁne-tuned ResNet50 obtained

89.95% of balanced accuracy on the validation set, as

shown in Table 2. We concluded that the ﬁne-tuning

Figure 6: Confusion matrix of the predictions on the test

set.

technique is paramount to adapt the network trained

over a general problem domain to the Retinal OCT

images domain.

3.3.3 Test Set Evaluation

Driven by the results on the validation set, we evaluate

the ﬁne-tuned ResNet50 model on the test set, achiev-

ing a balanced accuracy of 98.04%. We conclude that

our method can generalize the knowledge learned on

the training set to new and unseen images.

Additionally, we generated the confusion matrix

of the predictions on the test set and present it in Fig-

ure 6, which indicates that our method achieves more

than 90% of accuracy in each class. Also, our method

had 0% of false positive and false negative classiﬁ-

cations for the Normal class, i.e., none of the patients

with some disease (CNV, DME, or Drusen) were clas-

siﬁed as Normal, nor healthy patients were misdiag-

nosed with CNV, DMR or Drusen. This is particularly

important considering a triage scenario, in which an

automatic model would decide if a patient needs to

be analyzed by expert clinicians or not. However, the

classiﬁcation of the Drusen class had 5.8% of mis-

ICAART 2022 - 14th International Conference on Agents and Artiﬁcial Intelligence

578

takes for CNV class, showing that further investiga-

tions can be done to mitigate these wrong predictions.

3.3.4 Interpreting Model Decisions

While our goal is to correctly distinguish among the

existing diseases present in OCT scans, developing

models with resilient and transparent decision rules is

also paramount. In this vein, we use Grad-CAM (Sel-

varaju et al., 2017), an AI explaining technique based

on Class Activation Maps, to highlight the most con-

tributing regions considered by the softmax classiﬁer

in its decision process. Figure 7 illustrates the class-

based salient regions for multiple images in our test-

ing set.

We observe that the model focuses on reasonable

regions of interest when classifying samples belong-

ing to the classes CNV, DME, and Normal. On the

other hand, it has incorrectly focused on artiﬁcial cap-

turing properties (e.g., the presence of image borders

created by incorrect placement of the OCT sample

during capturing) on two degenerated cases of the

class Drusen. This indicates that the model tends to

use characteristics unrelated to the disease itself to

classify a scan. This goes in accordance with the re-

sults presented in Figure 6, whose errors might be a

reﬂex of the model incorrectly identifying unrelated

artifacts in the image. Besides that, these cases show

that additional preprocessing and regularization steps

that remove such capturing artifacts might lead to a

more robust model that correctly focuses on the dis-

criminative features of Drusen.

4 CONCLUSION

The use of deep learning to classify medical images

proved to be effective in multiple medical problems,

achieving high accuracy and potentially alleviating

the need for manual inspection. In practice, however,

automatic approaches are still far from being success-

fully deployed in most real-world scenarios. Deep

neural networks are complex non-linear models prone

to capture biases and ﬂaws in the dataset, exploiting

them during training but failing to generalize to un-

seen data.

This work discusses potential biases that may

be introduced during dataset collection, organization,

and model training. To exemplify them, we evaluated

convolutional neural networks for retinal OCT image

classiﬁcation of eye diseases — Choroidal Neovascu-

larization, Diabetic Macular Edema, and Drusen. We

employed transfer learning for different CNN archi-

tectures, training a fully-connected layer on top of the

CNV

Figure 7: Class-based saliency maps for samples in the test

set. Samples are labeled as, from the left-most to the right-

most columns: CVN, DME, Drusen and Normal.

extracted features. The best architecture, a ResNet50,

achieved a balanced accuracy of 79.75% on the vali-

dation set. In an additional experiment, we ﬁne-tuned

it, allowing all network weights to be freely updated

for this task, which considerably improved the re-

sults to 89.95% on the same set. When evaluating

the test set, our method obtained a balanced accuracy

of 98.04%.

Interpretability experiments highlighted that the

model correctly considers relevant retinal regions for

most classes. However, for Drusen samples, the

model exploits artifacts on the image border instead

of focusing on discriminative portions of the retina.

These are common telltales that the network has in-

correctly captured a potential bias (in this case, due

to the noise in training images) that would probably

affect its performance in unseen Drusen imagery. Be-

sides that, such behavior indicates that a more rigor-

ous capturing procedure and preprocessing step could

improve model robustness and conﬁdence for the im-

plementation in real-world scenarios.

Bias Assessment in Medical Imaging Analysis: A Case Study on Retinal OCT Image Classiﬁcation

579

In future work, we will investigate further the pos-

sible biases present in medical imaging analysis prob-

lems. We will extend the evaluation to other datasets,

employing interpretability techniques to aid us in cat-

egorizing existing biases in data. Additionally, we

will investigate which preprocessing techniques are

viable to reduce the impact of noise and acquisition

artifacts of images on model performance.

ACKNOWLEDGMENTS

This research was supported by S

ao Paulo Research

Foundation (FAPESP) [grant numbers 2015/11937-

9, 2017/12646-3 and 2017/21957-2], and the Na-

tional Council for Scientiﬁc and Technological De-

velopment (CNPq) [grant numbers 140929/2021-5,

161015/2021-2 and 304380/2018-0].

REFERENCES

Deepak, S. and Ameer, P. (2019). Brain tumor classiﬁcation

using deep cnn features via transfer learning. Comput-

ers in Biology and Medicine, 111:103345.

Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., and Fei-Fei,

L. (2009). ImageNet: A large-scale hierarchical image

database. In IEEE International Conference on Com-

puter Vision and Pattern Recognition (CVPR), pages

248–255.

He, K., Zhang, X., Ren, S., and Sun, J. (2016). Deep resid-

ual learning for image recognition. In IEEE Inter-

national Conference on Computer Vision and Pattern

Recognition (CVPR), pages 770–778.

Howard, A. G., Zhu, M., Chen, B., Kalenichenko,

D., Wang, W., Weyand, T., Andreetto, M., and

Adam, H. (2017). MobileNets: Efﬁcient convolu-

tional neural networks for mobile vision applications.

arXiv:1704.04861.

Hussain, M., Bird, J. J., and Faria, D. R. (2018). A Study on

CNN Transfer Learning for Image Classiﬁcation. In

UK Workshop on Computational Intelligence (UKCI),

pages 191–202. Springer.

Kermany, D., Zhang, K., and Goldbaum, M. (2018a).

Labeled Optical Coherence Tomography (OCT) and

Chest X-Ray Images for Classiﬁcation. Mendeley

Data, 2(2).

Kermany, D. S., Goldbaum, M., Cai, W., Valentim, C. C.,

Liang, H., Baxter, S. L., McKeown, A., Yang, G., Wu,

X., Yan, F., Dong, J. D., Prasadha, M. K., Pei, J., Ting,

M. Y., Zhu, J., Li, C., Hewett, S., Dong, J., Ziyar, I.,

Shi, A., Zhang, R., Zheng, L., Hou, R., Shi, W., Fu,

X., Duan, Y., Huu, V. A., Wen, C., Zhang, E. D. Z.,

Zhang, C. L., Li, O., Wang, X., Singer, M. A., Sun, X.,

Xu, J., Tafreshi, A., Lewis, M. A., Xia, H., and Zhang,

K. (2018b). Identifying Medical Diagnoses and Treat-

able Diseases by Image-Based Deep Learning. Cell,

172(5):1122–1131.

Litjens, G., Kooi, T., Bejnordi, B. E., Setio, A. A. A.,

Ciompi, F., Ghafoorian, M., Van Der Laak, J. A.,

Van Ginneken, B., and S

anchez, C. I. (2017). A survey

on deep learning in medical image analysis. Medical

Image Analysis, 42:60–88.

Oliveira, G., Padilha, R., Dorte, A., Cereda, L., Miyazaki,

L., Lopes, M., and Dias, Z. (2020). COVID-19 X-

ray Image Diagnostic with Deep Neural Networks. In

2020 Brazilian Symposium on Bioinformatics (BSB),

pages 57–68. Springer.

Roberts, M., Driggs, D., Thorpe, M., Gilbey, J., Yeung,

M., Ursprung, S., Aviles-Rivero, A. I., Etmann, C.,

McCague, C., Beer, L., et al. (2021). Common pitfalls

and recommendations for using machine learning to

detect and prognosticate for covid-19 using chest ra-

diographs and ct scans. Nature Machine Intelligence,

3(3):199–217.

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

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

explanations from deep networks via gradient-based

localization. In IEEE International Conference on

Computer Vision (ICCV), pages 618–626.

Shi, F., Wang, J., Shi, J., Wu, Z., Wang, Q., Tang, Z., He, K.,

Shi, Y., and Shen, D. (2020). Review of artiﬁcial in-

telligence techniques in imaging data acquisition, seg-

mentation, and diagnosis for COVID-19. IEEE Re-

views in Biomedical Engineering, 14:4–15.

Simonyan, K. and Zisserman, A. (2014). Very deep con-

volutional networks for large-scale image recognition.

arXiv:1409.1556, pages 1–14.

Swanson, E. A. and Fujimoto, J. G. (2017). The ecosystem

that powered the translation of oct from fundamental

research to clinical and commercial impact. Biomedi-

cal Optics Express, 8(3):1638–1664.

Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., and Wojna,

Z. (2016). Rethinking the inception architecture for

computer vision. In IEEE International Conference

on Computer Vision and Pattern Recognition (CVPR),

pages 2818–2826.

Tan, C., Sun, F., Kong, T., Zhang, W., Yang, C., and Liu, C.

(2018). A Survey on Deep Transfer Learning. In In-

ternational Conference on Artiﬁcial Neural Networks

(ICANN), pages 270–279. Springer.

Tan, M. and Le, Q. (2019). Efﬁcientnet: Rethinking model

scaling for convolutional neural networks. In IEEE In-

ternational Conference on Machine Learning (ICML),

pages 6105–6114.

Wu, J., Ruan, S., Lian, C., Mutic, S., Anastasio, M. A.,

and Li, H. (2018). Active learning with noise mod-

eling for medical image annotation. In 15th Inter-

national Symposium on Biomedical Imaging (ISBI),

pages 298–301. IEEE.

Yosinski, J., Clune, J., Bengio, Y., and Lipson, H. (2014).

How transferable are features in deep neural net-

works? In Advances in Neural Information Process-

ing Systems (NIPS), pages 3320–3328.

Zhuang, F., Qi, Z., Duan, K., Xi, D., Zhu, Y., Zhu, H.,

Xiong, H., and He, Q. (2021). A Comprehensive Sur-

vey on Transfer Learning. Proceedings of the IEEE,

109(1):43–76.

ICAART 2022 - 14th International Conference on Agents and Artiﬁcial Intelligence

580