PROJECT REPORT

Benign and Malignant Colon Cancer Prediction using Deep CNN models

Name: Nithish S
Reg. No: 953621243039
Company Name: Hera Diagnostics, Rajapalayam
Time Frame: 26/06/2023 to 04/08/2023

1. Project Description
Usually while diagnosing a patient with cancer,the diagnosis process involves multiple
steps such as non-invasive for initial screening and invasive procedures after reliable diagnosis
from initial screening. The whole process of both stages is time consuming and mostly the
invasive procedures such as biopsies (removing a piece of muscle/tissue/fluid/blood around the
cancer affected region mostly through surgery) consumes a lot of time. After removing a sample
from the patient, the sample is placed in a slide and gets stained using chemicals such as
hematoxylin and eosin and they are left to dry. Then the slides are manually analyzed by a
pathologist which takes hours. A pathologist usually looks for the features like Cell Polarity,
Cell Size and Shape, Invasion into Stroma, Mitotic Figures, Chromatin Content and Nuclear
Cytoplasmic Ratio.
● Problem Statement: The main purpose of the project is to reduce the efforts of the
pathologist who spends hours analyzing the slide samples.
● Goals: Developing an AI model which augments the process of diagnosis and not fully
taking control over the pathologist's job.
● Dataset: The dataset used is an open source dataset which contains only images and has
no annotations. Each image has a dimension of 768x768 pixels. The data is divided into
two categories 1.Benign and 2. Adenocarcinoma, both parts are equally divided and we
have 5000 images in each category.

2. Methodology
 We can use AI methods such as CNN(Convolutional Neural Network) based algorithms
for precise classification of cells. These Algorithms can do the next level of learning from
images by extracting features in front of the images and identify the patterns underlying in the
input images.In addition to CNN-based algorithms, implementing transfer learning with pre-
trained models like VGG16 or ResNet can expedite the training process and enhance the model's
accuracy by leveraging features learned from large image datasets. This combination of cutting-
edge CNN architectures and transfer learning will provide a robust foundation for our AI-
assisted cell classification system.
Day wise Tasks Completed during the internship period

Day Date Task Done

1 26/06/2023 • Project kickoff involves officially starting the project,

discussing goals, and clarifying the scope.

• Define the objectives clearly, including what the AI

model should achieve.

• Understanding the dataset requirements entails identifying

the dataset source, format, size, and any specific
preprocessing needs.

2 27/06/2023 • Collecting and organizing the dataset involves acquiring

the dataset you'll be working with.

• Ensure that you have a balanced distribution of benign

and adenocarcinoma images to prevent bias in the model.

• Organize the data in a structured manner for ease of

access during later stages.

3 28/06/2023 • On this day, I wrote code or used image processing

libraries (such as OpenCV or PIL) to resize each image to a
predefined dimension, such as 224x224 pixels.

• Ensure that the aspect ratio is maintained during resizing

to prevent distortion.

• This consistency in image dimensions will make it easier

to feed the data into your neural network during training.

4 30/06/2023 • Complete the resizing of all images and perform a quality

check and Continue resizing any remaining images from
the dataset.
 • After resizing, visually inspect a sample of resized images
to ensure they look appropriate and haven't been distorted.

• Additionally, check the file format (e.g., JPEG, PNG) to

ensure consistency.

5 01/06/2023 • I began applying data augmentation techniques to

increase the dataset size.

• Data augmentation involved applying various

transformations to the existing images to create new
training examples. Common augmentation techniques
included Rotations, Flips, Translations, Brightness and
Contrast Adjustments, and Zoom.

• I implemented these techniques using libraries like

Augmentor, OpenCV, or image augmentation functions
available in deep learning frameworks like TensorFlow and
PyTorch. Data augmentation helped the model generalize
better by exposing it to a wider range of variations in the
training data.

6 03/07/2023 • Continuing with data augmentation, I ensured that a

sufficient number of augmented images were generated to
expand the training dataset.

• I kept track of the transformations applied to each image

for later reference.

• Additionally, I started documenting my data

preprocessing procedures, as this documentation would be
valuable for reproducibility and future reference.

7 04/07/2023 • Splitting the dataset into training and testing sets ensures
that you have a separate portion of data for training and
evaluating the model.

• A common split is 80% for training and 20% for testing to

assess model performance.

• After Augmentation and Splitting the DataSet for training

and testing, healthcheck on the dataset was performed
which ensures proper training of the CNN model.

8 05/07/2023 Read Research Papers about different CNN architectures

that are available, and analyze the advantages and
disadvantages of each architecture.
 After getting to know multiple CNN architectures, I
understood that a proper architecture is required for better
feature extraction and better confidence of model.
Taking these into consideration, I designed a simple CNN
architecture.

9 06/07/2023 • The focus is on establishing the initial structure of the

deep learning model.

• This involves defining the neural network architecture,

selecting appropriate activation functions, specifying the
loss function relevant to the classification task, and
deciding on weight initialization methods.

• Choices regarding the architecture might include the

number and type of layers, whether to employ well-
established architectures like VGG16 or ResNet, or
designing a custom architecture tailored to the dataset.

10 07/07/2023 • On this day attention shifts to the configuration of the

training pipeline.

• This encompasses critical decisions such as optimizer

selection (e.g., Adam, SGD, RMSprop), setting an initial
learning rate, determining batch size for mini-batch
training, and considering regularization techniques to
mitigate overfitting.

• These two days(9 & 10) lay the foundation for subsequent
model development and training phases, ensuring a well-
structured and optimized neural network ready for learning
from the data.

11 08/07/2023 • On Day 11, I embarked on the crucial task of model

development. This phase involved crafting the neural
network architecture, whether it was a predefined
Convolutional Neural Network (CNN) or a custom design
tailored to my dataset.

• I determined the arrangement of layers, including

convolutional, pooling, and fully connected layers, as well
as the activation functions used between them, such as
Rectified Linear Units (ReLU).

• This architecture formed the backbone of my deep

learning model and was essential for feature extraction
from input images, setting the stage for accurate cell
 classification in subsequent phases of the project.

12 10/07/2023 • On Day 12, I translated my chosen neural network

architecture into code, configuring the layers, activations,
and other architectural elements.
• This involved setting up the model's structure, specifying
input and output dimensions, and compiling the model with
the chosen loss function and optimizer.

• Subsequently, I trained the model on a limited portion of

my dataset, typically a smaller subset. This initial testing
allowed for a rapid assessment of the model's behavior,
helping me catch any glaring issues or architectural flaws
early on.

13 11/07/2023 • On Day 13, I reviewed the results from Day 12's initial
testing. If the model exhibited any shortcomings, this was
the time to make necessary adjustments to its architecture,
hyperparameters, or preprocessing steps.

• Fine-tuning involved experimenting with different

learning rates, batch sizes, or regularization techniques.

• After these refinements, I validated the model on a more

substantial subset of the dataset, providing a more
representative assessment of its performance and
generalization capabilities.

14 12/07/2023 • On Day 14, I continued refining the model iteratively,

considering the insights gained from validation and
performance on the smaller dataset.

• I focused on achieving a balance between model

complexity and generalization. Additionally, I started
calculating and documenting key performance metrics,
such as accuracy, precision, recall, and F1-score.

• These metrics provided a quantitative measure of the

model's effectiveness in classifying cells, guiding further
adjustments if necessary.

15 13/07/2023 • On this day, I focused on analyzing the results obtained

from the initial training of my model.

• This crucial step involved a comprehensive examination

of various aspects, including loss curves, accuracy, and the
model's behavior on the validation data.

• By closely scrutinizing these metrics, I gained valuable

 insights into how well my model was performing and
where it might need improvement.
• The loss curves provided information about the
convergence of the training process, while accuracy metrics
offered a glimpse into the model's classification
capabilities.

• The behavior on validation data allowed me to assess the

model's generalization. This analysis laid the foundation for
informed decision-making regarding the model's
refinement.

16 14/07/2023 • Building on the insights gained from Day 15's analysis,

Day 16 was dedicated to implementing optimizations to
enhance the model's performance.

• This encompassed a range of actions, such as fine-tuning

hyperparameters to strike the right balance between
underfitting and overfitting, introducing regularization
techniques like dropout or L2 regularization to combat
overfitting, and potentially making adjustments to the
model architecture itself.

• These optimizations aimed to address the identified areas

for improvement and refine the model's capabilities. By
iteratively refining the model based on data-driven insights,
I ensured that it continued to evolve and meet the
performance standards required for my specific
classification task.

17 15/07/2023 • On Day 17, I embarked on the pivotal step of model

training. This involved feeding the entire training dataset
into my neural network and iteratively adjusting the
model's parameters to minimize the chosen loss function.

• Training a deep learning model, especially a complex one,

often demanded significant computational resources, which
included GPU acceleration.

• The duration of this process varied, and it could take

several hours or even days, contingent on factors like
model complexity and dataset size. This phase allowed the
model to uncover intricate patterns and relationships within
the data.

18 17/07/2023 • On Day 18, the model training continued from where it

left off. It was essential to monitor the training process
closely, keeping an eye on key indicators like training loss
 and accuracy.

• I also considered early stopping strategies to prevent

overfitting, wherein training stopped when validation
performance ceased to improve.

• The goal was to ensure that the model converged to a

point where it had learned to generalize effectively from
the training data.

19 18/07/2023 • On this day, I delved into fine-tuning, a critical aspect of

model refinement.

• Fine-tuning involved making iterative adjustments to the

model architecture, hyperparameters, and other factors
based on observations and insights gained from the
validation results during the earlier stages of training.

• This step aimed to enhance the model's performance,

increase its ability to generalize, and address any potential
issues or limitations identified.

20 19/07/2023 • On Day 20, fine-tuning efforts persist. The adjustments

made to the model continue to be refined based on insights
and observations.

• Concurrently, validation on the validation dataset is

carried out to gauge the impact of these adjustments on the
model's performance.

• This iterative process of fine-tuning ensures that the

model reaches an optimal state, achieving the highest
possible accuracy and robustness while minimizing
overfitting.

21 20/07/2023 • On Day 21, evaluating the model's performance on the

test dataset is a pivotal step in assessing its real-world
applicability.

• This process involves exposing the model to new, unseen

data, mirroring its performance in practical scenarios.

• It serves as a litmus test, revealing how well the model

generalizes beyond the training data and highlighting any
potential issues such as overfitting or underfitting.

22 21/07/2023 • Moving to Day 22, the focus turns to quantifying the

 model's performance comprehensively. Calculating and
documenting key performance metrics, including accuracy,
precision, recall, and F1-score, offers an objective and
clear understanding of the model's effectiveness in
classification tasks.

• These metrics go beyond a mere accuracy percentage,

providing insights into the model's ability to correctly
classify various classes, identify false positives, and assess
its overall precision and recall.

• This documentation becomes instrumental in gauging the

model's suitability for its intended medical diagnosis
application and guiding any necessary refinements.

23 22/07/2023 • On Day 23, I focused on creating a user-friendly

graphical user interface (GUI) using Flask, HTML, CSS,
and JavaScript.

• I designed the layout, styled elements with CSS, and

added interactivity with JavaScript.

• This phase aimed to establish a visually appealing and

responsive interface for pathologists to interact with the AI
model.

24 24/07/2023 • Continuing from Day 23, Day 24 involved integrating the

GUI with Flask for backend functionality.

• User inputs were processed, and real-time results from

the AI model were provided.

• Thorough testing and user feedback guided refinements

in HTML, CSS, and JavaScript, ensuring a smooth and
effective user experience for pathologists interacting with
the AI model through the web-based GUI

25 25/07/2023 • On Day 25, the focus was on integrating the AI model

with the pathologist's workflow by connecting the
graphical user interface (GUI) with the model.

• This involved ensuring a smooth and secure data flow

between the GUI and the AI model, allowing pathologists
to use the tool effectively.

26 26/07/2023 • Day 26 was dedicated to rigorous testing and validation

of the integrated system. Comprehensive testing scenarios
included user interactions, data inputs, and system
 responses.
• This ensured a reliable and stable system, free from
usability or technical issues, and validated the accuracy of
AI model predictions within the GUI context.

27 27/07/2023 • On Day 27, I documented the entire project, including

code, model architecture, usage instructions, and
troubleshooting steps.

• This documentation was critical for future reference,

ensuring that both users and developers could understand
and utilize the system effectively.

• The emphasis was on creating well-organized and

comprehensive documentation.

28 28/07/2023 • Day 28 was dedicated to reviewing and finalizing project

documentation to ensure it was comprehensive, accurate,
and user-friendly.

• I paid close attention to clarity, organization, and

accessibility for individuals with varying levels of
technical expertise.

• This step guaranteed that the project's knowledge was

well-preserved and readily accessible for all stakeholders.

29 31/07/2023 • On Day 29, I conducted thorough final testing and

addressed any remaining issues or glitches to ensure that
the project was well-prepared for deployment in a real-
world setting.

• This phase involved subjecting the entire system,

including the AI model, GUI, and integration components,
to comprehensive testing scenarios.

• The goal was to identify and rectify any anomalies,

usability issues, or technical glitches that could hinder the
smooth operation of the system.

30 01/08/2023 • On Day 30, the focus remained on the optimization of the

AI model. I delved into fine-tuning model parameters and
hyperparameters, leveraging user feedback and
performance evaluations.

• This involved adjusting learning rates, batch sizes, or

regularization techniques to enhance the model's overall
performance. The aim was to ensure that the AI model
operated at its peak accuracy and robustness levels.
 31 02/08/2023, • Continuing into Day 31, the emphasis shifted towards
03/08/2023 refining the user interface (UI) based on user feedback and
usability testing.

• I worked on making the UI more intuitive, visually

appealing, and user-friendly. This included revising layout
designs, improving user navigation, and optimizing the
overall user experience.

• By aligning the UI more closely with user expectations

and needs, I ensured that the interaction between
pathologists and the AI-assisted system was as smooth and
effective as possible.

32 04/08/2023 • On the final day, Day 32, I prepared for a project

presentation to my internship team or supervisor.

• The presentation was compelling and covered key aspects

of the project, including its objectives, development
process, achieved results, and future potential applications.

• I was ready to field questions and provide insights into

my AI-assisted cell classification system, highlighting its
significance and the value it brought to the medical
diagnosis domain.

• This presentation marked the culmination of my efforts

and showcased the project's readiness for real-world
implementation and further advancements.

Conclusion
Through this project, we have successfully achieved several significant goals aimed at enhancing
the diagnostic process in the field of cell classification for cancer diagnosis. Our primary
objective was to develop an AI-assisted cell classification system that augments the pathologist's
workflow, reducing their workload while maintaining a high standard of accuracy and reliability.

One of the key accomplishments was the development of a robust deep learning model, based on
Convolutional Neural Networks (CNN), which demonstrated exceptional capabilities in
accurately classifying cells into benign and adenocarcinoma categories. This model, fine-tuned
through rigorous optimization, significantly reduces the time and effort required for manual slide
analysis. It acts as a valuable second opinion for pathologists, aiding in the identification of
crucial features such as cell polarity, size, shape, invasion into stroma, mitotic figures, chromatin
content, and nuclear-cytoplasmic ratio

.
1. Adenocarcinoma Samples 2. Benign Samples

Furthermore, the integration of this AI model into a user-friendly graphical user interface (GUI)
marked another milestone. The GUI, built using Flask, HTML, CSS, and JavaScript, ensures
seamless interaction between pathologists and the AI system. It provides a platform where users
can effortlessly upload cell images, receive prompt analysis, and access classification results in a
comprehensible format.

3.Training and Validation Loss 4. Training and Validation Accuracy

In addition to the technical achievements, the project prioritized usability and user feedback.
Continuous optimization based on user inputs led to an interface that not only streamlines the
diagnostic process but also caters to the specific needs and expectations of medical professionals.
This collaborative approach ensured that the system aligns with the highest standards of medical
accuracy and usability.
 5. Confusion Matrix

As a result of these accomplishments, we have created an AI-assisted cell classification system

that significantly reduces the efforts of pathologists without compromising accuracy. It
demonstrates the potential to revolutionize the diagnostic process, offering a powerful tool for
early cancer detection and efficient disease management.

6. Final Review From Project Team

In conclusion, this project underscores the successful fusion of cutting-edge AI technology with
medical expertise. It represents a significant step forward in the quest to enhance cancer
diagnosis, reduce human error, and ultimately improve patient outcomes. Our achievement lies
not only in the development of a sophisticated AI model and user-friendly interface but also in
the promise of a brighter future for medical professionals and patients alike.