Comparative Analysis of Learning Strategies for Multi-Magnification

Pathological Image Classification

Yixuan Pu

School of Electrical and Computer Engineering, The University of Sydney, New South Wales, Australia

Keywords: Colorectal Histopathology, Multi-Channel, Stepwise Learning.

Abstract: The automatic classification of pathology images plays a crucial role in computer-aided diagnosis by

enhancing diagnostic efficiency and minimizing human error. In this paper, the Enteroscope Biopsy

Histopathological H&E Image Dataset (EBHI) is utilized to systematically compare and analyze the

performance of three strategies—Single-Magnification Training, Multi-Channel Fusion, and Stepwise

Cumulative Learning—to optimize pathology image classification. The Single-Magnification Training

strategy serves as a baseline experiment to validate the optimization effect of the model, achieving the highest

classification accuracy of 94.64% at 200× magnification. Under strict filtering conditions, Multi-Channel

Fusion achieves a peak classification accuracy of 96.06%. However, this approach remains inferior to

Stepwise Cumulative Learning. This learning strategy significantly outperforms training solely at the highest

magnification, achieving a classification accuracy of 98.27% on 400× images. This study demonstrates that

the cumulative learning strategy effectively enhances the classification performance of pathology images.

Low-magnification images contribute to improving the classification accuracy of high-magnification images,

offering new insights into multi-scale feature fusion, dynamic learning strategies, and computer-aided

pathology diagnosis. Furthermore, this study validates the applicability of the EBHI dataset in multi-

magnification pathology analysis and advances the development of intelligent pathology image analysis.

1 INTRODUCTION

Early diagnosis of colorectal cancer is crucial for

reducing its high morbidity and mortality, with

pathological image analysis remaining the gold

standard for diagnosis. Traditional pathology analysis

relies on the manual evaluation of tissue sections by

pathologists, a process that is not only time-

consuming and labor-intensive but also prone to

subjective bias. Consequently, leveraging advanced

technologies to enhance the efficiency and accuracy

of pathological image analysis has become a focal

point in contemporary medical research.

In recent years, deep learning techniques have

provided an efficient and accurate solution for the

automatic classification and detection of pathology

images by automatically extracting image features.

Current research focuses on the following three key

areas: optimizing deep learning models, improving

data preprocessing and handling class imbalance, and

integrating multi-magnification information.

https://orcid.org/0009-0003-9188-1252

Firstly, in terms of model optimization, numerous

studies have sought to enhance the accuracy of

pathology image classification by refining deep

learning architectures. Khan et al. (2024) proposed a

Swin Transformer-based approach that leverages the

self-attention mechanism for feature extraction in

pathology images. By incorporating normalized

preprocessing, their method significantly improves

classification accuracy. Kim et al. (2021)

systematically compared the performance of

Convolutional Neural Network (CNN), Residual

Neural Network (ResNet), and Vision Transformer

(ViT) in pathology image classification. However,

these studies primarily focused on optimizing a single

model and did not integrate multi-magnification

information.

Secondly, in terms of data preprocessing and class

imbalance handling, studies have demonstrated that

appropriate data augmentation and class balancing

strategies can enhance classification performance.

Malik et al. (2019) investigated the impact of various

222

Pu, Y.

Comparative Analysis of Learning Strategies for Multi-Magniﬁcation Pathological Image Classiﬁcation.

DOI: 10.5220/0013681400004670

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

In Proceedings of the 2nd International Conference on Data Science and Engineering (ICDSE 2025), pages 222-229

ISBN: 978-989-758-765-8

preprocessing methods on pathology image

classification and found that data augmentation

improves the model’s generalization ability. Raju and

Rao (2022) addressed the issue of class imbalance by

proposing a deep learning framework that integrates

Class-Balanced Loss with data augmentation,

enabling the model to better recognize minority class

samples. Although these approaches improve model

stability to some extent, most existing methods are

designed for single-magnification images and fail to

fully exploit the potential complementary information

across different magnifications.

Finally, in the fusion of multi-magnification

information, researchers have explored various

approaches to integrating different magnifications to

enhance the classification accuracy of pathology

images. Das et al. (2017) employed Majority Voting

Fusion by independently training a CNN model at

different magnifications and fusing classification

results from multiple perspectives at the inference

stage, thereby improving the overall classification

accuracy of full-slide images. However, this method

treats different magnifications as independent

information sources and fails to fully model the

hierarchical relationship between them, making it less

effective in simulating the gradual magnification

process that pathologists naturally follow in real-

world diagnosis.

This study aims to develop and validate an

efficient pathology image classification model based

on deep learning, explore optimization strategies for

different magnifications, and conduct an in-depth

investigation into data missing issues, magnification

combination methods, and multi-magnification

learning. The main innovations of this study include:

1) Validate and improve the single-magnification

training method and establish baseline

experiments;

2) Investigate the applicability of the 12-channel

fusion model and compare different data

imputation strategies;

3) Propose a cumulative magnification learning

strategy based on a progressive training

sequence.

2 MATERIAL AND METHODS

2.1 Dataset

The EBHI dataset, developed through a collaboration

between Northeastern University and China Medical

University Cancer Hospital, serves as a standardized

dataset for the automated classification of colorectal

cancer histopathological images (Hu et al., 2023). The

dataset comprises 5,532 electron microscope images,

categorized into five pathological groups: Normal,

Polyp, Low-grade IN, High-grade IN, and

Adenocarcinoma. Among these, the first two classes

represent non-cancerous conditions, while the latter

three exhibit pre-cancerous or malignant

characteristics to varying degrees. In this study, a

classification task was constructed based on the

distinction between benign and malignant tissue types,

following this categorization criterion.

In the data preprocessing stage, standardization

was applied to the raw data to ensure stable model

training. Size normalization was first performed to

adapt input images to the required format for deep

learning models and to maintain compatibility with

commonly used pre-trained CNN architectures

(Tellez et al., 2019). Specifically, all pathological

images were uniformly resized from 2048 × 1536 to

224 × 224 to ensure a consistent input size.

Additionally, to enhance training stability and

accelerate convergence, pixel value normalization

was conducted, scaling all pixel intensities to the

range [0,1] to minimize the impact of numerical

differences between images on model training.

Following this, all images were converted to

Tensor format to enable efficient batch processing in

PyTorch. To further enhance data diversity, data

augmentation techniques were applied to the training

set, including random horizontal flip, color jitter, and

random cropping. These augmentation strategies

were introduced to improve the model’s

generalization ability, allowing it to better adapt to

pathological images under varying conditions (Hao et

al., 2021; Tellez et al., 2019; Yuan, 2021).

In terms of data partitioning, all images were

divided into 40% for the training set, 40% for the

validation set, and 20% for the test set, ensuring that

all magnification images from the same case appeared

in only one of these subsets to prevent data leakage

(Hu et al., 2023).

Moreover, since some cases contained multiple

images at a specific magnification, a magnification

combination sampling strategy was employed. This

approach involved randomly combining different

images from the same case to expand the dataset and

enhance the robustness of the model (Hashimoto et

al., 2020; Tokunaga et al., 2019).

2.2 Methodology

This study employs Residual Network-50 (ResNet50)

as the fundamental deep-learning model for the

classification of colorectal cancer histopathological

images. To evaluate the impact of different training

Comparative Analysis of Learning Strategies for Multi-Magniﬁcation Pathological Image Classiﬁcation

223

strategies, three experimental schemes were designed

and implemented: Single-Magnification Training,

Multi-Channel Fusion, and Stepwise Cumulative

Learning. A systematic analysis was conducted to

assess the classification performance of each strategy.

To ensure the comparability of experiments and

control variables, ResNet50 was consistently used in

all experiments, with modifications made to its input

layer (conv1) based on specific experimental

requirements.

ResNet50 is a deep convolutional neural network

(CNN) based on the Residual Network architecture. It

incorporates residual block structures and skip

connections to effectively mitigate the vanishing

gradient problem in deep networks (He et al., 2016).

This network has strong feature extraction

capabilities, enabling it to learn deep structural

information from pathological images. To

accommodate different input strategies, two

configurations of the ResNet50 input layer were

implemented in this study: For Single-Magnification

Training, the input channel was set to 3 channels

(standard RGB structure), ensuring that the model

learns pathological features at a single magnification.

For Multi-Channel Fusion, the input channel was

adjusted to 12 channels (4 magnifications × 3 RGB

channels), allowing the model to integrate multi-

magnification information within a single input image

and simultaneously learn structural features across

different magnifications.

To systematically investigate the impact of

different magnification levels on model classification

performance, this study designed two major

experimental schemes.

The first scheme, Single-Magnification Training,

involved independently training the model using

images at 40×, 100×, 200×, and 400× magnifications

to analyze classification performance at each

magnification level.

The second scheme, Multi-Magnification

Learning, included two distinct strategies: Multi-

Channel Fusion and Stepwise Cumulative Learning.

In the Multi-Channel Fusion experiment, 40×, 100×,

200×, and 400× magnification images were

concatenated following the RGB structure, forming a

12-channel input, which was then trained using

ResNet50 to evaluate the effect of cross-

magnification information fusion.

In the Stepwise Cumulative Learning experiment,

the model was progressively trained by sequentially

incorporating magnification information in the order

of 40× → 40×+100× → 40×+100×+200× →

40×+100×+200×+400× to examine whether gradual

learning enhances the generalization capability of the

model. At each training stage, testing was conducted

at the highest magnification level learned up to that

point (e.g., after training on 40×+100×+200×, the

final evaluation was performed on 200×) to assess

whether low-magnification information contributes to

improving classification performance at higher

magnifications.

Furthermore, considering that some cases may

lack corresponding images at certain magnifications,

three different data processing strategies were

designed in the Multi-Channel Fusion experiments to

investigate the impact of different filling methods on

model performance. These three strategies are Strict

Filtering, Black Filling, and Nearest Magnification

Filling.

2.3 Evaluation Metrics

This study employs Accuracy, Precision, Recall,

Specificity, and F1-Score as evaluation metrics to

comprehensively assess the model’s performance in

classifying Benign and Malignant tissues. Accuracy,

which measures the overall correctness of

classifications, is widely used in decision-making

models (Turing, 2009). Precision, reflecting the

reliability of malignant predictions, is a critical metric

in medical image analysis (Van Rijsbergen, 1979).

Recall evaluates the model’s ability to detect

malignant cases, while Specificity assesses its

capability to distinguish between benign and

malignant tissues (Altman & Bland, 1994). F1-Score,

as the harmonic mean of Precision and Recall, is

particularly useful for handling imbalanced datasets

(Van Rijsbergen, 1979).

3 RESULTS AND DISCUSSION

3.1 Single-Magnification Training

In this study, single-magnification training was first

conducted on pathological images at different

magnifications to validate model optimization and

investigate the impact of magnification levels on

classification performance.

Due to significant differences in tissue structural

information and cellular feature representation across

different magnifications, their performance in

classification tasks also varies. Low-magnification

images provide an overview of the tissue structure,

whereas high-magnification images reveal more

detailed cellular features. These differences influence

the classification performance at different

magnification levels. Therefore, experiments were

ICDSE 2025 - The International Conference on Data Science and Engineering

224

conducted by independently training ResNet50 at 40×,

100×, 200×, and 400× magnifications, and their

classification performance was compared.

Table 1. Classification Performance of Single-Magnification Training

nification Accurac

Cate

Precision Recall S

ecificit

F1-score

40× 90.67

Beni

n 90.27 94.20 85.71 92.19

Mali

nant 91.30 85.71 94.20 88.42

100× 92.18

Benign 88.76 98.13 87.42 93.21

Malignant 97.88 87.42 98.13 92.35

200× 94.64

Benign 92.28 94.10 94.98 93.18

Mali

nant 96.19 94.98 94.10 95.58

400× 94.72

Beni

n 96.51 94.04 95.60 95.25

Malignant 92.55 95.60 94.04 94.05

Table 1 presents the classification performance of

the ResNet50 model at different magnifications (40×,

100×, 200×, and 400×). As shown in the table,

Classification Accuracy exhibits an overall increasing

trend with higher magnifications, reaching 94.64% at

200× and further improving to 94.72% at 400×.

Additionally, for the malignant category, both

Precision and Recall at 200× and 400× magnifications

are higher than those at lower magnifications,

indicating that high-magnification images are more

beneficial for malignant lesion detection.

This study utilizes the same EBHI dataset and

adopts ResNet50 as the baseline model, consistent

with the original study. By implementing a series of

data preprocessing and training optimization

strategies, the classification accuracy of the model has

been significantly improved. Compared to the highest

classification accuracy of 83.81% reported in the

original study using ResNet50, the optimized

strategies in this study have achieved 94.64%

accuracy at 200× magnification, demonstrating a

remarkable performance enhancement.

In the Single-Magnification Training experiment,

this study employed a data augmentation strategy to

increase data diversity and enhance the model’s

generalization capability. All pathological images

were normalized to the range [0,1] and resized to 224

× 224 for input. The data augmentation operations

included random horizontal flipping, vertical flipping,

90°, 180°, and 270° rotations, as well as color

jittering, enabling the network to develop greater

robustness to rotational transformations.

In terms of training optimization, this study

employed a dynamic learning rate adjustment method

(ReduceLROnPlateau) to adapt the learning rate at

different training stages, thereby preventing

convergence issues that may arise from a fixed

learning rate. Additionally, the Adam optimizer was

used in place of traditional Stochastic Gradient

Descent (SGD), leveraging momentum and adaptive

learning rate mechanisms to enhance training stability

and accelerate convergence.

3.2 Multi-Channel Fusion

The results of single-magnification training indicate

that images at different magnifications exhibit

varying classification performances. Among them,

high-magnification images at 200× and 400×

achieved better classification accuracy, though the

performance improvement between these two

magnifications was relatively minor. This

phenomenon suggests that relying solely on a single

magnification may not fully capture the

discriminative features of pathological images.

To address this, the study further investigates

whether the fusion of multi-magnification

information can enhance classification performance.

Compared to Single-Magnification Training, Multi-

Channel Fusion integrates information from multiple

scales, enabling the model to learn both macro-level

tissue structures and fine-grained cellular morphology

simultaneously. This approach improves

classification robustness and generalization

capability. Therefore, this study explores a Multi-

Channel Fusion strategy, in which images of the same

lesion at different magnifications are concatenated

into a 12-channel input, enhancing the model’s ability

to learn across different scales and ultimately

improving classification performance.

In the experimental design, each case contains

images at four magnifications (40×, 100×, 200×, and

400×), which are concatenated into a unified input

and fed into a modified ResNet50 model for training.

However, in the dataset, some cases lack images at

certain magnifications. To investigate the impact of

different missing data handling strategies on model

performance, this study explores the following three

Comparative Analysis of Learning Strategies for Multi-Magniﬁcation Pathological Image Classiﬁcation

225

approaches:

1) Strict Filtering: Cases missing images at any

magnification are directly excluded, ensuring

that both the training and testing samples are

complete 12-channel inputs.

2) Black Filling: If an image at a specific

magnification is missing, it is replaced with a

black image (all pixel values set to 0) to maintain

input size consistency.

3) Nearest Magnification Filling: If an image at a

particular magnification is missing, it is replaced

by an image from the nearest lower

magnification. For example, if the 40× image is

missing, the 100× image is used instead; if 100×

is also missing, the 200× image is used, and so

on.

All models are trained on preprocessed multi-

channel images, and testing is also performed on

concatenated 12-channel inputs, rather than

evaluating single-magnification images separately.

By comparing the effects of these three data-filling

strategies, this study analyzes how different missing

data-handling approaches influence classification

accuracy.

Table 2. Comparison of Classification Performance Across Multi-Channel Fusion Strategies

Strategy Accuracy Category Precision Recall Specificity F1-

score

Strict Filter 96.06

Benign 96.28 95.58 96.52 95.92

Mali

nant 95.86 96.52 95.58 96.19

Black Filling 95.41

Beni

n 94.27 96.72 94.10 95.48

Mali

nant 96.61 94.10 96.72 95.34

Nearest Magnification Filling 92.96

Benign 95.71 90.74 95.45 93.16

Malignant 90.21 95.45 90.74 92.76

Table 2 presents the impact of three data filling

strategies on the multi-channel fusion classification

task. The experimental results show that the Strict

Filtering strategy achieved the highest classification

accuracy of 96.06%, indicating that complete multi-

magnification information provides the most stable

feature representation, thereby enhancing the model’s

classification capability.

In contrast, when using Black Filling to replace

missing magnification images, classification

performance declined but still maintained a relatively

high accuracy. This suggests that the model can

partially adapt to the Black Filling strategy; however,

zero-value images can introduce additional noise,

leading to less stable feature representation compared

to complete data.

Nevertheless, the Black Filling strategy

outperformed the Nearest Magnification Filling

approach, suggesting that maintaining data

consistency is more beneficial than filling missing

magnifications with available images from other

magnifications.

3.3 Stepwise Cumulative Learning

Although the multi-magnification fusion strategy

improves classification performance by integrating

images from different magnifications, it does not fully

reflect the observation sequence followed by

pathologists during actual diagnosis. Typically,

pathologists begin with low-magnification images to

examine the overall tissue structure before

progressively zooming in to higher magnifications to

obtain finer details. Simply relying on multi-channel

concatenation may not fully leverage these

hierarchical features.

To address this, this study proposes a stepwise

training strategy, namely Stepwise Cumulative

Learning, in which the model is initially trained on

low-magnification images and then progressively

incorporates higher-magnification information. This

approach aims to investigate whether the gradual

accumulation of magnification information can

enhance the final classification performance. To

validate the effectiveness of stepwise learning, the

model is evaluated during the testing phase only on

the highest magnification introduced in the training

process.

During training, the model initially uses only 40×

magnification images for preliminary training,

allowing it to learn global tissue structure information

at a low magnification. Subsequently, images at 100×,

200×, and 400× magnifications are progressively

introduced, simulating a stepwise optimization

process supported by multi-scale information.

Notably, at each training stage, all case IDs must

remain consistent, ensuring that images of the same

case across different magnifications originate from

the same source. This constraint prevents the model

from merely learning single-magnification features

and instead enables it to establish robust cross-

magnification associations.

ICDSE 2025 - The International Conference on Data Science and Engineering

226

By following this design, the model not only

leverages the holistic structural information provided

by low magnifications but also gradually integrates

fine-grained details from higher magnifications,

ultimately improving classification performance.

Table 3. Comparison of Classification Performance in the Stepwise Cumulative Learning Strategy

nification Accurac

Cate

Precision Recall S

ecificit

F1-score

40× 90.67

Beni

n 90.27 94.20 85.71 92.19

Mali

nant 91.30 85.71 94.20 88.42

40×+100× 95.43

Benign 97.45 94.26 96.91 95.83

Malignant 93.06 96.91 94.26 94.94

40×+100×+200× 96.09

Benign 95.00 95.52 96.49 95.26

Mali

nant 96.87 96.49 95.52 96.68

40×+100×+200×+400× 98.27

Beni

n 99.06 97.69 98.94 98.37

Malignant 97.39 98.94 97.69 98.16

Table 3 presents the classification performance of

the Stepwise Cumulative Learning strategy as

different magnification images are progressively

incorporated. The experiment begins with training

exclusively on 40× magnification images, followed

by the sequential addition of 100×, 200×, and 400×

magnifications, allowing the model to gradually

establish connections between global low-

magnification structural information and high-

magnification fine details.

The results demonstrate that the Stepwise

Cumulative Learning strategy consistently

outperforms single-magnification training across all

high-magnification testing tasks. For instance, in the

400× magnification task, the classification accuracy

of Single-Magnification training was 94.72%,

whereas Stepwise Cumulative Learning

(40×→100×→200×→400×) improved the accuracy

to 98.27%, achieving a 3.55% increase. Similarly, in

the 200× magnification task, Single-Magnification

training achieved an accuracy of 94.64%, while

Stepwise Cumulative Learning (40×→100×→200×)

improved the accuracy to 97.31%, representing a

2.67% increase.

These findings indicate that introducing low-

magnification information helps enhance the model’s

classification capability, and as higher magnification

information is progressively accumulated during

training, the overall model performance is further

optimized.

A more detailed analysis of classification

performance reveals that the Stepwise Cumulative

Learning strategy provides the most significant

improvement in the detection of Malignant cases. For

instance, in the 400× magnification task, the Recall

for malignant cases increased from 94.04% in Single-

Magnification training to 98.94%, representing a

4.9% improvement. This suggests that stepwise

learning helps the model better capture malignant

lesion characteristics.

In contrast, the improvement in classification

precision for Benign cases was relatively smaller. For

example, in the 400× magnification task, the

Precision increased only slightly from 96.51% to

97.39%. This discrepancy may be attributed to the

greater complexity of malignant pathological

features, where the stepwise learning strategy

provides richer hierarchical information, enabling the

model to identify malignant patterns more accurately.

Compared to the Multi-Channel Fusion strategy,

the primary advantage of Stepwise Cumulative

Learning is its alignment with the observation

sequence used by pathologists. By maintaining case

ID consistency and emphasizing the sequential

progression of magnification, this approach enables

the model to leverage low-magnification information

to refine high-magnification classification.

In summary, the Stepwise Cumulative Learning

strategy outperforms both Single-Magnification

Training and Multi-Channel Fusion across all high-

magnification testing tasks, with the most significant

improvement observed in malignant case

identification. These findings suggest that a

progressive learning approach incorporating low-

magnification information is a crucial method for

enhancing the classification accuracy of pathological

images.

3.4 Discussion

This study explores the impact of different

pathological image magnifications on classification

performance by employing three training strategies.

However, certain limitations remain, which should be

addressed in future research.

First, this study employs ResNet50 as the sole

baseline model. Although it has demonstrated strong

performance in pathological image classification, its

Comparative Analysis of Learning Strategies for Multi-Magniﬁcation Pathological Image Classiﬁcation

227

reliance on local receptive fields may limit its ability

to integrate cross-magnification information

effectively. Future research could explore self-

attention-based architectures, such as ViT or Swin

Transformer, to improve global feature extraction.

Additionally, leveraging DenseNet or other feature-

reuse networks may enhance robustness, especially in

small-sample scenarios.

Second, although the EBHI dataset was used to

validate the proposed approach, further experiments

on larger and more diverse pathological datasets are

necessary to assess the generalizability and stability

of the method. Additionally, in real-world clinical

practice, pathologists rely not only on static images

but also on clinical history and lesion evolution over

time. Future research should explore Multimodal

Fusion Models, integrating multi-magnification

information with other clinical data, to enhance

diagnostic decision-making.

Overall, while this study demonstrates the

effectiveness of stepwise learning and multi-

magnification fusion, further improvements in model

selection, dataset diversity, and clinical applicability

are necessary to enhance the practical deployment of

such methods in pathology.

4 CONCLUSION

This study systematically investigates the impact of

multi-magnification information on pathological

image classification by designing and validating three

learning strategies: Single-Magnification Training,

Multi-Channel Fusion, and Stepwise Cumulative

Learning. The experiments, conducted using

ResNet50 on the EBHI dataset, demonstrate the

effectiveness of the proposed strategies in enhancing

classification performance.

The results confirm that the proposed strategies

significantly enhance classification performance. In

Single-Magnification Training, the classification

accuracy was improved from the previously reported

highest accuracy of 83.81% to 94.64% at 200×

magnification through the optimization techniques

applied in this study. Stepwise Cumulative Learning

achieved the highest accuracy among all strategies,

particularly in malignant pathology detection, where

it further improved classification accuracy to 98.27%

on 400× test images. Additionally, the study

highlights the impact of different missing

magnification image filling strategies, showing that

the Strict Filtering approach yields the best

classification performance (96.06%).

These findings suggest that progressively

incorporating low-magnification information

enhances the model’s ability to extract discriminative

features, improving overall classification accuracy.

Moreover, this study validates the suitability of the

EBHI dataset for multi-magnification learning

research, providing a useful reference for future

dataset selection.

In summary, this study presents a novel

optimization approach for multi-magnification

pathological image classification, laying the

groundwork for future advancements in intelligent

pathology image analysis.

REFERENCES

Altman, D. G., & Bland, J. M., 1994. Diagnostic tests. 1:

Sensitivity and specificity. BMJ: British Medical

Journal, 308(6943), 1552.

Das, K., Karri, S. P. K., Roy, A. G., Chatterjee, J., & Sheet,

D., 2017. Classifying histopathology whole-slides

using fusion of decisions from deep convolutional

network on a collection of random multi-views at multi-

magnification. In 2017 IEEE 14th International

Symposium on Biomedical Imaging (ISBI 2017), pp.

1024-1027. IEEE.

Hao, R., Namdar, K., Liu, L., Haider, M. A., & Khalvati, F.,

2021. A comprehensive study of data augmentation

strategies for prostate cancer detection in diffusion-

weighted MRI using convolutional neural networks.

Journal of Digital Imaging, 34, 862-876.

Hashimoto, N., Fukushima, D., Koga, R., Takagi, Y., Ko,

K., Kohno, K., et al., 2020. Multi-scale domain-

adversarial multiple-instance CNN for cancer subtype

classification with unannotated histopathological

images. In Proceedings of the IEEE/CVF conference on

computer vision and pattern recognition, pp. 3852-

3861.

He, K., Zhang, X., Ren, S., & Sun, J., 2016. Deep residual

learning for image recognition. In Proceedings of the

IEEE conference on computer vision and pattern

recognition, pp. 770-778.

Hu, W., Li, C., Rahaman, M. M., Chen, H., Liu, W., Yao,

Y., et al., 2023. EBHI: A new Enteroscope Biopsy

Histopathological H&E Image Dataset for image

classification evaluation. Physica Medica, 107, 102534.

Khan, A. A., Arslan, M., Tanzil, A., Bhatty, R. A., Khalid,

M. A. U., & Khan, A. H., 2024. Classification of colon

cancer using deep learning techniques on

histopathological images. Migration Letters, 21(S11),

449-463.

Kim, S. H., Koh, H. M., & Lee, B. D., 2021. Classification

of colorectal cancer in histological images using deep

neural networks: An investigation. Multimedia Tools

and Applications, 80(28), 35941-35953.

Malik, J., Kiranyaz, S., Kunhoth, S., Ince, T., Al-Maadeed,

S., Hamila, R., & Gabbouj, M., 2019. Colorectal cancer

ICDSE 2025 - The International Conference on Data Science and Engineering

228

diagnosis from histology images: A comparative study.

arXiv preprint arXiv:1903.11210.

Raju, M. S. N., & Rao, B. S., 2022. Colorectal multi-class

image classification using deep learning models.

Bulletin of Electrical Engineering and Informatics,

11(1), 195-200.

Tellez, D., Litjens, G., Bándi, P., Bulten, W., Bokhorst, J.

M., Ciompi, F., & Van Der Laak, J., 2019. Quantifying

the effects of data augmentation and stain color

normalization in convolutional neural networks for

computational pathology. Medical image analysis, 58,

101544.

Tokunaga, H., Teramoto, Y., Yoshizawa, A., & Bise, R.,

2019. Adaptive weighting multi-field-of-view CNN for

semantic segmentation in pathology. In Proceedings of

the IEEE/CVF conference on computer vision and

pattern recognition, pp. 12597-12606.

Turing, A. M., 2009. Computing machinery and

intelligence, pp. 23-65. Springer Netherlands.

Van Rijsbergen, C. J., 1979. Information Retrieval, 2nd edn.

Newton, MA.

Yuan, Z., 2021. Deep Learning Based Data Augmentation

for Breast Lesion Detection. Doctoral dissertation,

Carnegie Mellon University Pittsburgh, PA.

Comparative Analysis of Learning Strategies for Multi-Magniﬁcation Pathological Image Classiﬁcation

229