Live Cell Stage Classification Using Deep Learning

Uma Mahesh R N, Kushal S M, Ponnanna K V, Sanjan B M and Vishnu S

Dept of CSE (AI&ML), ATME College of Engineering, Mysore, Karnataka, India

Keywords: Live Cell Imaging, Deep Learning, Sequential Convolutional Neural Network (SCNN), ResNet50,

EfficientNetB0, Cell Stage Classification, Interphase, Mitosis, Data Enrichment, Compound Scaling,

Receiver Operating Characteristic (ROC), Confusion Matrix, Transfer Learning, Imagenet Pretraining,

Explainable AI, Biomedical Applications, Cellular Behavior Studies.

Abstract: Live cell imaging has transformed biological research, offering real-time insight into dynamic cellular

processes. This project focuses on using deep learning techniques to automate the detection and

classification of live cell stages, specifically distinguishing between the interphase and mitosis phases.

Traditional methods, such as fluorescence microscopy and flow cytometry, are highly dependent on manual

or semiautomated, time-intensive and error-prone approaches. Our proposed solution employs advanced

deep learning architectures, including Sequential Convolutional Neural Network (SCNN), ResNet50, and

EfficientNetB0, to overcome these limitations. The data set used comprises high-resolution images of

nematode cells, preprocessed using resizing, normalization, and data augmentation techniques to ensure

robust model training. The performance of each model is evaluated on the basis of metrics such as accuracy,

positive predictive value (PPV), sensitivity, and the F1 score. In particular, EfficientNetB0 emerges as the

model with the best performance, achieving a test accuracy of 98%, showcasing its superior ability to

generalize in diverse data.

1 INTRODUCTION

Live cell imaging has transformed biological

research by enabling real-time observation of

cellular processes such as mitosis and signal

transduction. Despite its advantages, manual

analysis of live cell imaging data is time-consuming

and error-prone, especially in distinguishing phases

such as Interphase and Mitosis. This project

addresses these challenges using deep learning

methods, specifically a Sequential Convolutional

Neural Network (SCNN) and pretrained models like

ResNet50 and EfficientNetB0. These models

leverage compound scaling and transfer learning

from ImageNet to achieve high accuracy in binary

classification tasks. By automating live cell stage

classification, this project accelerates cellular

analysis, offering a scalable and efficient solution

for research and medical diagnostics, with

significant implications for cancer studies and drug

discovery.

1.1 Objective

The primary objective of this project is to develop an

automated system for live cell stage classification,

specifically distinguishing between Interphase and

Mitosis phases, to reduce reliance on manual

methods. State-of-the-art deep learning

architectures, including SCNN, ResNet50, and

EfficientNetB0, are utilized to ensure high accuracy

and efficiency. Data preprocessing techniques such

as resizing, normalization, and augmentation (e.g.,

rotations, flips, and brightness adjustments) are

implemented to enhance model robustness and

generalization. Model performance is evaluated

using metrics like accuracy, positive predictive value

(PPV), sensitivity, and F1-score, and ROC to ensure

consistent and reliable results. The system aims to

support biomedical research by enabling scalable

applications such as cellular behavior analysis,

cancer stage identification, and drug testing.

414

Mahesh R N, U., S M, K., K V, P., B M, S. and S, V.

Live Cell Stage Classiﬁcation Using Deep Learning.

DOI: 10.5220/0013593300004664

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

In Proceedings of the 3rd International Conference on Futuristic Technology (INCOFT 2025) - Volume 2, pages 414-421

ISBN: 978-989-758-763-4

1.2 Literature Survey

The classification of cell cycle stages using deep

learning has been an active area of research. Several

studies have explored various methods and datasets

to enhance accuracy. Below is a summary of key

related works:

Robust Classification of Cell Cycle Phase and

Biological Feature Extraction by Image-Based Deep

Learning: Okada et al. (2020) proposed a method

using convolutional neural network (CNN) to

classify fluorescence images of cells into G1/S and

G2 phases without relying on specific cell cycle

markers. The study achieved an accuracy of

approximately 90%. Using Grad-CAM analysis, the

authors identified critical subcellular features that

contributed to the classification decisions. (Nagao,

Sakamoto, et al. , 2020).

Cell Cycle Stage Classification Using Phase

Imaging with Computational Specificity (PICS):

Nguyen et al. (2022) introduced a label-free deep

learning method for classifying cell cycle stages

based on single-shot quantitative Live Cell Stage

Classification Using Deep Learning phase imaging.

Their model achieved comparable accuracy to

traditional techniques, with atleast one stage in

interphase classification below 95% accuracy. (He,

Kandel, et al. , 2022).

Predicting Cell Cycle Stage from 3D Single-Cell

Nuclear-Stained Images: Li et al. (2024) applied a

CNN-based model to classify cell cycle stages using

3D nuclear-stained single-cell images. Their model

achieved an accuracy of 93%, showcasing the

potential of combining 3D imaging and deep

learning. (Li, Nichols, et al. , 2024).

Cell Cycle Classification Using Imaging Flow

Cytometry and Deep Learning: Zhang et al. (2022)

developed deep learning models, including a 2-layer

fully connected neural network, to classify cell cycle

stages from imaging flow cytometry data. Despite

exploring various architectures, the best-balanced

accuracy achieved was below 95%. This study

indicated room for improvement in both model

design and preprocessing techniques when using

imaging flow cytometry data for cell stage

classification (Rade, Zhang, et al. , 2022).

Deep Learning-Based Reconstruction of

Embryonic Cell-Division Cycle in Nematodes:

Wang et al. (2024) focused on the classification of

cell division stages in nematode embryos using

multiple CNN architectures. The models achieved

accuracies below 95%, highlighting the difficulties

associated with embryonic cell cycle stage

classification. This research pointed out the

challenges of dealing with complex and dynamic

datasets, particularly in embryonic imaging (Wang

et al. , 2024).

Each of these studies contributes valuable insights to

the field of cell cycle classification using deep

learning. However, the reported accuracies below

95% indicate significant opportunities for

improvement. The current project aims to build upon

these works by leveraging advanced architectures,

robust preprocessing techniques, and optimized

training methods to achieve higher accuracy and

scalability.

2 DESIGN AND PRINCIPLE OF

OPERATION

2.1 Proposed System

2.1.1 Data Preprocessing

The system begins with data preprocessing to en-

sure high-quality inputs for the models. The dataset

comprises high-resolution images of nematode cells

labeled as Interphase or Mitosis. Each image is

resized to 224 × 224 pixels and normalized to the

range [0, 1]. To enhance model robustness and

prevent over-fitting, data augmentation techniques

such as random rotations, flips, and brightness

adjustments are applied, ensuring the models

generalize effectively to unseen data.

2.1.2 Model Architectures

The proposed system employs three deep learning

models: SCNN, ResNet50, and EfficientNetB0. The

SCNN is a custom-built architecture that uses

convolutional layers for feature extraction, max-

pooling layers for dimensionality reduction, and

fully connected dense layers with dropout to

mitigate overfitting. ResNet50, pretrained on the

ImageNet dataset, is fine-tuned for binary

classification by replacing the final layers with task-

specific dense layers, leveraging residual learning to

address the vanishing gradient problem.

EfficientNetB0, known for its compound scaling,

balances network depth, width, and resolution,

making it both accurate and computationally

efficient. This model is fine-tuned for the current

application and achieves the best performance

among the three.

Live Cell Stage Classiﬁcation Using Deep Learning

415

2.1.3 Training and Optimization

The models are trained using the Adam optimizer

with a dynamic learning rate scheduler, which

adjusts the learning rate during training for better

convergence. Binary cross-entropy loss is employed

as it is well-suited for binary classification tasks.

Throughout the training process, metrics such as

accuracy, loss, positive predictive value (PPV),

sensitivity, and F1-score are monitored to ensure

convergence and prevent overfitting.

2.1.4 Evaluation Metrics

The system’s performance is evaluated using a

variety of metrics. Accuracy measures the overall

correctness of the model, while positive predictive

value (PPV) and sensitivity quantify its ability to

correctly classify positive cases and retrieve all

relevant instances. The F1-score provides a balance

between positive predictive value (PPV) and

sensitivity. A confusion matrix visualizes

classification performance across the two classes,

and a Receiver Operating Characteristic (ROC)

analyzes the trade-off between sensitivity and

specificity, further validating the model’s reliability.

2.1.5 System Workflow

The system workflow begins with preprocessing the

input dataset, followed by training and fine-tuning

the three models. During the evaluation phase, the

models’ performance metrics are analyzed, and the

best-performing model, EfficientNetB0, is selected

for deployment. The system outputs the classified

cell stage (Interphase or Mitosis) with high

confidence.

2.2 Flow Chart of the Proposed System

Figure 1: Flow Chart of Proposed System

2.3 Methodology

The methodology for this project involves a systematic

approach to classify live cell stages, focusing on

Interphase and Mitosis phases. The process is divided into

several key steps: dataset preparation, preprocessing,

model architecture, training, and evaluation.

2.3.1 Dataset and Preprocessing

The dataset consists of high-resolution images of

nematode cells, labeled as either Interphase or

Mitosis. To ensure consistency, all images are

resized to 224 × 224 pixels and normalized to a

range of [0, 1]. This preprocessing step standardizes

the input for all models, enabling efficient training

and reducing computational overhead. Data

augmentation techniques are applied to improve

generalization and prevent overfitting. These

techniques include:

1. Random rotations to simulate various

orientations of cells.

2. Horizontal and vertical flips to account for

variability in image orientation.

3. Brightness adjustments to simulate

different imaging conditions.

2.3.2 Model Architectures

Three deep learning models are employed for

this task: Sequential Convolutional Neural Network

(SCNN), ResNet50, and EfficientNetB0. Each

model architecture is optimized to achieve high

accuracy and efficiency.

1) Sequential Convolutional Neural

Network (SCNN): The SCNN is a custom-

built model tailored for this application. It

consists of:

• Multiple convolutional layers for

feature extraction.

• Max-pooling layers to reduce spatial

dimensions and computational

complexity.

• Fully connected dense layers for

classification.

• Dropout layers to prevent overfitting

during training.

INCOFT 2025 - International Conference on Futuristic Technology

416

Figure 2: Architecture of Sequential CNN

2) ResNet50: ResNet50, a well-established

model pretrained on ImageNet, is utilized

for its residual learning capabilities. The

key features include:

• Residual blocks that mitigate the

vanishing gradient problem.

• Pretrained weights from ImageNet,

fine-tuned for binary classification.

• A final dense layer customized for the

classification of Interphase and

Mitosis.

Figure 3: Architecture of ResNet50 model used in the

proposed system.

3) EfficientNetB0: EfficientNetB0 is chosen

for its compound scaling capabilities,

optimizing depth, width, and resolution for

maximum accuracy and computational

efficiency. Its key features include:

• Balanced architecture using compound

scaling for resource optimization.

• Pretrained on ImageNet and fine-tuned

for this application.

• Superior generalization capabilities,

making it the best-performing model in

this study.

Figure 4: Architecture of EfficientNetB0 model.

2.3.3 Training and Optimization

The models are trained using the following settings:

• Optimizer: Adam optimizer with an initial

learning rate of 0.001.

• Loss Function: Binary cross-entropy loss,

suitable for binary classification tasks.

• Batch Size: 32 images per batch for

balanced training.

• Epochs: Models are trained for up to 50

epochs, with early stopping based on

validation accuracy.

• Learning Rate Scheduler: Dynamically

adjusts the learning rate to optimize

convergence.

2.3.4 Evaluation Metrics

The performance of the models is evaluated

using the following metrics:

• Accuracy: Measures the overall

correctness of predictions.

• Positive Predictive Value (PPV):

Calculates the ratio of label true positive

(TP) predictions to total predicted positives.

• Sensitivity: Measures the model’s ability to

identify all relevant instances (true

positives (TP)).

• F1-Score: Provides a balance between

positive predictive value (PPV) and

sensitivity.

• Confusion Matrix: Visualizes the

classification performance for each class.

• Receiver Operating Characteristic

(ROC): Evaluates the trade-off between

sensitivity and specificity.

2.3.5 System Workflow

The workflow of the proposed methodology is out-

lined as follows:

1) Preprocess the dataset by resizing,

normalizing, and augmenting the images.

2) Train the three models (SCNN, ResNet50,

and EfficientNetB0) using the prepared

dataset.

3) Evaluate model performance on test data

using the defined metrics.

4) Select the best-performing model,

Efficient-NetB0, for deployment based on

accuracy, positive predictive value (PPV),

sensitivity, and F1-score.

Live Cell Stage Classiﬁcation Using Deep Learning

417

2.3.6 Model Architectures

• Sequential CNN: Features are extracted

using multiple Conv2D and MaxPooling2D

layers, followed by fully connected Dense

layers.

• ResNet50: Pretrained on ImageNet, it

utilizes residual learning for feature

extraction. The final layers are customized

for binary classification.

• EfficientNetB0: With compound scaling,

this model optimizes accuracy and

computational efficiency.

2.4 Implementation

The implementation of this project involves de-

signing and training deep learning models to classify

live cell stages. The following steps outline the com-

plete implementation process, from data preparation

to model evaluation:

2.4.1 Data Preparation

The dataset consists of labeled high-resolution

images of nematode cells categorized as Interphase

or Mitosis. The preprocessing pipeline includes:

• Resizing: All images are resized to 224 ×

224 pixels to maintain uniformity across

the dataset.

• Normalization: Pixel values are scaled to a

range of [0, 1] to ensure faster and more

stable convergence during training.

• Data Augmentation: Techniques such as

random rotations, horizontal and vertical

flips, and bright- ness adjustments are

applied to enhance dataset variability and

prevent overfitting.

2.4.2 Model Training

Three deep learning models—Sequential

Convolutional Neural Network (SCNN), ResNet50,

and EfficientNetB0—were implemented and trained

using TensorFlow for the binary classification of

Interphase and Mitosis stages. The SCNN was

custom-built with convolutional, max-pooling,

dense, and dropout layers to extract features and

prevent overfitting. ResNet50 and EfficientNetB0,

pretrained on ImageNet, were finetuned for the task

with their final layers replaced by task-specific

dense layers. Training utilized the Adam optimizer

with an initial learning rate of 0.001, binary cross-

entropy loss, and a batch size of 32 for upto 50

epochs, with early stopping based on validation loss

to avoid overfitting. A learning rate scheduler was

employed to ensure optimal convergence.

2.4.3 Evaluation Pipeline

The trained models were evaluated using a test

set. Various metrics were calculated to assess the

performance of each model:

Accuracy: The ratio of correctly predicted

instances to the total number of instances.

Positive predictive value (PPV): The

proportion of label true positive (TP) predictions

among all positive predictions.

Sensitivity: The proportion of label true positive

(TP) correctly identified out of all actual positives.

F1-Score: The harmonic mean of positive

predictive value (PPV) and sensitivity, providing a

balanced evaluation metric.

Confusion Matrix: A detailed breakdown of

label true positives (TP), true negatives (TN), false

positives (FP), and false negatives (FN).

Receiver Operating Characteristic (ROC): A

graphical representation of the trade-off between

sensitivity and specificity.

2.4.4 Implementation Workflow:

The complete workflow of the implementation is as

follows:

• Dataset Preparation: Preprocessing and

augmenting the dataset to create a robust

input pipeline.

• Model Training: Training the SCNN,

ResNet50, and EfficientNetB0 models on

the preprocessed dataset.

• Performance Evaluation: Using the

evaluation pipeline to compute metrics for

each model.

• Model Selection: Selecting EfficientNetB0

as the best-performing model based on its

superior accuracy of 98%.

• Deployment: Preparing the final trained

EfficientNetB0 model for integration into

biomedical research workflows.

2.4.5 Implementation Tools and

Environment

The following tools and libraries were used for

implementation:

• Programming Language: Python 3.8.

INCOFT 2025 - International Conference on Futuristic Technology

418

• Deep Learning Framework: TensorFlow

and Keras for model design, training, and

evaluation.

• Hardware: NVIDIA GPU for accelerated

training.

• Development Environment: Google

Colab and Jupyter Notebook for coding and

experimentation.

This implementation pipeline ensures a robust and

scalable system for automating live cell stage

classification while maintaining high accuracy and

computational efficiency.

3 SIMULATION RESULTS AND

ANALYSIS

This section presents the outcomes of the imple-

mented deep learning models for classifying live

cell stages into Interphase and Mitosis. The results

are evaluated using various metrics and

visualizations to demonstrate the performance of the

models and compare their effectiveness.

3.1 EfficientNetB0

The EfficientNetB0 model outperformed other

models with consistent performance across training,

validation, and test datasets:

• Training Accuracy: The model achieved

a high training accuracy of 98%.

• Validation Accuracy: A validation

accuracy of 99% demonstrated excellent

generalization.

• Test Accuracy: The test accuracy of 98%

confirmed the robustness of the model on

unseen data.

Performance Visualizations:

Figure 5. Loss and accuracy graphs on training and

validation sets for EfficientNetB0.

Confusion Matrix: The confusion matrix for

EfficientNetB0 highlights its strong classification

performance:

• True Positives (Mitosis): 361

• True Negatives (Interphase): 530

• False Positives: 13

• False Negatives: 4

Figure 6: Confusion Matrix for EfficientNetB0.

3.2 ResNet50

The ResNet50 model demonstrated competitive

performance but underperformed compared to

Efficient-NetB0:

• Training Accuracy: 90%.

• Validation Accuracy: 85%.

• Test Accuracy: 76%.

Performance Visualizations:

Figure 7: Loss and accuracy graphs on training and

validation sets for ResNet50.

Confusion Matrix: The confusion matrix for

ResNet50 revealed:

• True Positives (Mitosis): 534

• True Negatives (Interphase): 159

• False Positives: 0

• False Negatives: 215

Live Cell Stage Classiﬁcation Using Deep Learning

419

Figure 8: Confusion Matrix for ResNet50.

3.3 Sequential Convolutional Neural

Network (SCNN)

SCNN showed the lowest performance among the

three models:

• Training Accuracy: 76%.

• Validation Accuracy: 80%.

• Test Accuracy: 85%.

Performance Visualizations:

Figure 9: Loss and accuracy graphs on training and

validation sets for sequential convolutional neural network

(SCNN)

Confusion Matrix: The confusion matrix for SCNN

highlighted:

• True Positives (Mitosis): 506

• True Negatives (Interphase): 270

• False Positives: 28

• False Negatives: 104

Figure 10: Confusion Matrix for SCNN.

3.4 Comparison of Models

The three models were compared based on their

performance metrics:

• EfficientNetB0: Achieved the highest

accuracy and most stable performance

across all datasets, with minimal

fluctuations in validation accuracy and loss.

• ResNet50: Demonstrated moderate

performance, with occasional spikes in

validation loss and lower test accuracy.

• SCNN: Struggled with generalization and

stability, exhibiting fluctuations in

validation performance and relatively lower

test accuracy.

Figure 11: Comparison of Training and Validation Curves

for SCNN, ResNet50, and EfficientNetB0.

3.5 Key Findings

• EfficientNetB0: The best-performing model

with 98% test accuracy and superior

generalization, making it ideal for

deployment.

• ResNet50: While effective, it was less

stable and accurate compared to

EfficientNetB0.

INCOFT 2025 - International Conference on Futuristic Technology

420

• SCNN: Demonstrated limitations in

learning complex patterns, leading to lower

accuracy and inconsistent performance.

• EfficientNetB0 achieved superior accuracy

of 98%, significantly outperforming

ResNet50 and Sequential CNN. Table 1

summarizes the performance metrics.

Table 1: Performance Metrics of Models.

Model Accura

PPV Sensitivit

F1-

Score

SCNN 85% 82% 84% 83%

ResNet50 90% 88% 89% 88.5%

Efficient

Net

98% 97% 98% 97.5%

4 CONCLUSIONS

This project successfully demonstrates the use of

deep learning models for automating live cell stage

classification, focusing on Interphase and Mitosis.

Among the models evaluated, EfficientNetB0

achieved the highest performance with 98% test

accuracy, highlighting its superior generalization

and efficiency. The preprocessing techniques,

combined with metrics like accuracy, positive

predictive value (PPV), sensitivity, and confusion

matrices, ensured robust and reliable evaluations.

This system reduces manual effort and accelerates

cellular analysis, with potential applications in

cancer research, drug discovery, and biomedical

diagnostics. Future work will aim to extend

classification to all cell cycle stages and improve

model integration for real-world applications.

REFERENCES

Nagao, Y., Sakamoto, M., Chinen, T., Okada, Y. and

Takao, D.,2020. ”Robust classification of cell cycle

phase and biological feature extraction by image-

based deep learning”, Molecular biology of the cell,

31(13), pp.1346-1354.

He, Y.R., He, S., Kandel, M.E., Lee, Y.J., Hu, C., Sobh,

N., Anastasio, M.A. and Popescu, G., 2022. ”Cell

cycle stage classification using phase imaging with

computational specificity”, ACS photonics, 9(4),

pp.1264-1273.

Li, G., Nichols, E.K., Browning, V.E., Longhi, N.J.,

Camplis-son, C., Beliveau, B.J. and Noble, W.S.,

2024. ”Predicting cell cycle stage from 3D single-cell

nuclear-stained images”, bioRxiv.

Bernal, C.E., ”Cell Cycle Classification using Imaging

Flow Cytometry and Deep Learning”.

Khatri, D. and Athale, C.A., 2024. ”Deep learning-based

reconstruction of embryonic cell-division cycle from

label-free microscopy time-series of evolutionarily

diverse nematodes”, bioRxiv, pp.2024-05.

Rade, J., Zhang, J., Sarkar, S., Krishnamurthy, A., Ren, J.

and Sarkar, A., 2022. ”Deep learning for live cell

shape detection and automated afm navigation”,

Bioengineering, 9(10), p.522.

Rotman-Nativ, N. and Shaked, N.T., 2021. ”Live cancer

cell classification based on quantitative phase spatial

fluctuations and deep learning with a small training

set”, Frontiers in Physics, 9, p.754897.

Pattarone, G., Acion, L., Simian, M., Mertelsmann, R.,

Follo, M. and Iarussi, E., 2021. ”Learning deep

features for dead and living breast cancer cell

classification without staining”, Scientific reports,

11(1), p.10304.

Padovani, F., Mairh¨ormann, B., Falter-Braun, P.,

Lengefeld, J. and Schmoller, K.M., 2022.

”Segmentation, tracking and cell cycle analysis of

live-cell imaging data with Cell-ACDC”, BMC

biology, 20(1), p.174.

Gallusser, B., Stieber, M. and Weigert, M., 2023, October,

“Self-supervised dense representation learning for

live-cell microscopy with time arrow prediction”, In

International Conference on Medical Image

Computing and Computer-Assisted Intervention (pp.

537-547). Cham: Springer Nature Switzerland.

Bhandary, M., Reyes, J.P., Ertay, E. and Panda, A., 2022,

“Double U-Net for Super-Resolution and

Segmentation of Live Cell Images”, arXiv preprint

arXiv:2212.02028.

Jang, J., Lee, K. and Kim, T.K., 2023, “Unsupervised

Contour Tracking of Live Cells by Mechanical and

Cycle Consistency Losses”, In Proceedings of the

IEEE/CVF Conference on Computer Vision and

Pattern Recognition (pp. 227-236).

Mahesh, U. and Kiran, B., 2024, July. Three-dimensional

(3-D) objects classification by means of phase-only

digital holographic information using Alex Network.

In 2024 International Conference on Signal

Processing, Computation, Electronics, Power and

Telecommunication (IConSCEPT) (pp. 1-5). IEEE.

Live Cell Stage Classiﬁcation Using Deep Learning

421