Enhancing Blood Cancer Diognosis with Data Driven Techniques

Balaji J., Sanjai M., Hemadharshini V., Gunavathi R. and Gobisha K.

Department of Information Technology, Nandha College of Technology, Erode, Tamil Nadu, India

Keywords: Blood Cancer Classification, Deep Learning, resnet50 Model, Blood Cancer Detection.

Abstract: Blood is needed by the human body to carry essential nutrients, oxygen and immune cells. Leukemia and

lymphoma are two deliverables that can be traced to abnormalities in blood cells. Blood cancers which include

a range from leukemia to lymphomas and plasma cell diseases are among the most difficult to diagnose, yet

early detection is critical for effective treatment and positive clinical outcomes. In this scenario, the invention

of an integrated automated blood cancer detection system is essential. This pipeline classifies the blood sample

pictures and classifies between Benign, Early, Pre and Pro cancer types using a pre- trained ResNet50 model.

Images are pre-processed for it to comply with its input requirements, and then the model learns crucial

aspects to predict accurately. The model produced the type of cancer and a confidence score, which is an

indication of how likely the prediction is correct. The system developed using Keras as its deep learning

framework and Streamlit as its user interface, provides a reliable and portable tool that effectively automates

the picture processing procedure to aid with blood cancer identification. The implementation of these

consequences allows doctors in clinics to diagnose faster and more accurately by providing consistent and

accurate, classifications, while reducing manual work.

1 INTRODUCTION

Blood fills a major role in the human body,

delivering oxygen, nutrients, and immune cells

throughout organs and tissues. But in several blood

cancers, such as leukemia, lymphoma and myeloma,

mutations or uncontrolled growth in blood cells can

occur. Early detection of these tumors is needed for

effective treatment and improved survival rates.

Diagnosing blood cancer, a labor-intensive process

that is prone to human error, requires blood samples

to be painstakingly reviewed under microscope. The

rapid developments in the field of machine learning

and image analysis have made it increasingly

feasible to combine the automated identification and

categorization of blood malignancies.

By reducing the requirement for manual tasks and

facilitating better diagnostics, these technologies

could help with the fastest, efficient, and accurate

blood sample analysis. This method aims to assist

medical professionals by providing a more accurate,

more automated solution for detecting and classifying

blood malignancies, through the use of deep learning

models to classify images of blood samples. This will

eventually result in better diagnostic outcomes and

less costly treatments.

1.1 Blood Cancer Classification

Blood cancer classification is the process of

recognizing different types of blood cancer through

the study of blood characteristics. Blood

malignancies, including leukemia, lymphoma, and

myeloma, can have a myriad of presentations and

specific pathologic features. Classification is the

identification of those tumors which is based on the

abnormalities present in blood cells, including size,

shape, organization, etc. This step is critical in

determining the most appropriate treatment plan, as

well as assessing the severity of disease. Prompt and

accurate categorization allows healthcare

professionals to make informed decisions about the

most effective therapies, which greatly increases the

chance of well-timed success in any course of

therapy. Use of classification also ensures

identification of the stage or development of the

cancer which is important for long-term care and

judgment.

1.2 Deep Learning

Deep Learning is a powerful machine learning

technique back up by the architecture of multi-layered

738

J., B., M., S., V., H., R., G. and K., G.

Enhancing Blood Cancer Diognosis with Data Driven Techniques.

DOI: 10.5220/0013872000004919

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

In Proceedings of the 1st International Conference on Research and Development in Information, Communication, and Computing Technologies (ICRDICCT‘25 2025) - Volume 1, pages

738-743

ISBN: 978-989-758-777-1

neural networks to allowing system to learn by itself

and improve from data without given explicit

instruction. The features can be derived from

unprocessed input, for instance, text or pictures,

which eliminates the need for human feature

selection. Deep learning excels in uncovering

intricate interdependencies within volumes of data

that would inevitably elude conventional approaches.

As an example, for blood cancer diagnosis, deep

learning models such as convolutional neural

networks (CNNs) are allowed to learn how to analyze

images of blood samples and to detect subtle patterns

indicative of cancer. These models are trained on

high volumes of expert-labeled data and due to this,

they evolve over a period of time and can identify and

classify blood cancer types with high accuracy with

minimal human intervention.

1.3 Resnet50 Model

The ResNet50 Model is a well-known model type of

CNN a deep learning framework effective for

computer vision tasks, particularly in image

classification applications. Each layer of the model

has 50 layers to help it learn hierarchies of patterns,

allowing it to understand and capture complex

components in photographs. One of the major

aspects of ResNet50 is its use of residual connections,

which allow the network to reduce degradation in

performance as it deepens. Due to its ability to

maintain great accuracy despite having deep

architectures, ResNet50 is extremely useful for tasks

that involve large and complex datasets. The

ResNet50 model works exceptionally well in

classifying Such medical pictures including blood

samples by identifying the distinguishing

characteristics like size, shape, and texture of blood

cells.! Its ability to analyze huge amounts of data and

spot complex patterns makes it an ideal candidate for

automating blood cancer classification.

1.4 Blood Cancer Detection

Blood cancer detection focuses on studying abnormal

activity of blood cells in order to diagnosis of various

blood cancers. The process involves both traditional

diagnostic techniques such as manual inspection of

blood samples, as well as modern computer tools

such as deep learning for automatic segmentation.

The common way to detect blood cancer is through

finding deviations from normal blood cell properties,

such as atypical cell sizes, shapes or architectures that

may indicate the presence of malignancy.

Determining the most effective treatment strategies

and enhancing patient outcomes depend on the fast

and precise identification of blood malignancies. The

detection process may be improved and expedited

with the use of sophisticated models and algorithms,

giving medical practitioners quicker, more precise

findings that facilitate wise clinical judgment.

2 LITERATURE REVIEW

According to Md. Aslam Mollah et al, In order for a

sensor to be able to measure the salinity of seawater,

it has to have high sensitivity, structural simplicity,

and durability, according to the study published in

theJournal of Molecular Biology. In this work, an

ultrahigh sensitivity PCF salinity sensor was

proposed based on a sagnac interferometer (SI). The

finite element method (FEM) is used to investigate

the propagation properties of the suggested PCF. A

sensitivity of up to 37,500 nm/RIU and 7.5 nm/%

was achieved throughout the salinity of 0% to 100%

obtained. The PCF proposed here achieves the

highest resolutions of 2.66 × 10−06 RIU and 1.33 ×

10−02% as well as good linearities with the length of

2.20 cm is inferior to PCF with other lengths value,

up to 0.9924. Due to its remarkable results, this

proposed sensor has the ability to detect salinity in

seawater. Salt levels matter because they greatly

affect activities beneath the sea and ocean species.

But the whoifiée indicator of salinity is the electric

conductivity of the chlorite ions. But this

measurement is affected by the interference from

other contamination ions. Research has been attracted

due to the various benefits of fiber optic salinity

sensors such as their programmable birefringence,

small structure, remote sensing, tunable dispersion

and immunity to electromagnetic interference (EMI).

Here, a PCF salinity sensor based on the SI

phenomenon is studied numerically. Every one of the

PCF air gaps is believed to be filled with mixed salt

concentrations of the seawater (M.R.B.A. Faysal, et.,

al. 2020).

Most surface plasmon resonance (SPR)- based

photonic crystal fiber (PCF) sensors have been

reported for detecting the analyte refractive index

(RI) values between 1.33 and 1.41 (i.e.: human blood,

body fluid), whereas rising state-of the-art detection

techniques requires a more improved platform for the

accurate detection of HMIs. Here we propose a way

to mitigate this effect through a transformation of the

structure, which can be achieved using the existing

fabrication processes that the sensor architecture is

built from. We present our sensor that uses an air-

core PCF, and unlike existing PCF-based sensors,

Enhancing Blood Cancer Diognosis with Data Driven Techniques

739

the analyte enter the core of the PCF via vertical side

opening channel to measure RI of analytes higher

than that of the background of PCF. We employ a

chemically stable plasmonic (gold) material, and

since the plasmonic material is not directly contacted

by the analyte, the interference effect is minimized.

For the analyte RI of 1.42, the sensitivity, and

resolution of the spectrum have been found to be

11,700 nm/RIU and 8.55 × 10−6 RIU, respectively.

However, our proposed sensor still has the potential

to catch active samples of these chemical and

biological liquids. Various organizations have

released several surface plasmon resonance (SPR)

sensors in the past few years that detect analytes with

a range of 1.33 to 1.41. Micro- fluidic slotted sensors,

internal and external metal- coated PCF-based

sensors, nanowire-based sensors, and D-shaped

configuration-based sensors (A.K. Paul, et., al. 2020)

were derived from it, resulting in five types of PCF-

based SPR sensors.

Mohammad Al Mahfuz et al. In this work, we

have proposed a dual-core photonic crystal fiber (DC-

PCF)-based surface plasmon resonance (SPR)

biocompatible sensor for the refractive index (RI)

sensing of bio-organic molecules and biochemical

analytes in the visible to near-infrared (0.5 to 2 µm)

region. The sensor construction is easy with two

hexagonal ring lattices all with round air aperture.

The use of plasmonic material and an analyte

detecting layer on the outer surface of the fiber allows

practical applications to be made. Gold (Au) with 30

nm of thickness, a noble plasmonic material, is

utilized to excite the plasmons on the surface. It is

also suggested that a thin layer (~5 nm) of TiO2

(titanium oxide) acts as an interlayer cementing the

Au and the silica glass. Based on the mode solver the

finite element method (FEM) is utilized to investigate

the sensor response. Using both amplitude and

wavelength interrogation methods, numerical

findings reveal an optimal wavelength sensitivity

(WS) of as high as 28000 nm/RIU, an optimal

amplitude sensitivity (AS) of 6829 RIU−1, an optimal

amplitude resolution (AR) of 5x10−6 RIU, and a

wavelength resolution (WR) of 3.57x10−6 RIU for

the proposed sensor. Moreover, such a PCF-SPR

sensor, having 2800 RIU−1 as the highest FOMpeak

value, is currently the starkest sensor (M. Al Mahfuz,

et., al 2020).

In this work a novel technique named as 2-D

photonic crystal waveguid (PCW) based cell

detection has been proposed by Abinash Panda et al.

for identifying the nature of either normal or

malignant cells. The proposed metamaterial is

designed with 5 × 5 silicon-based rods on a square

lattice with a central defect and air as the background.

To correctly sense, we classify two sets of live cells:

Group I: Malignant Cell (YD-10B); Group II:

Normal Cell (INOK). Properly adjusting the plane

wave expansion (PWE) method, the electric field

distribution and the peak reflected wavelength have

been achieved in the designed PCW structure.

Accurate identification of normal and malignant cells,

a large number of structural parameters, including

but not limited to lattice spacing, circular rod

diameter, and backdrop material type, are needed.

This MATLAB simulation indicates that yellow

color (i.e. reflected wavelength) belongs to cancerous

cells and orange color (i.e. reflected wavelength)

corresponds to healthy cells. Similarly, the negative

dispersion coefficient, scattering loss, and the

nonlinear coefficient of the proposed structure are

precisely evaluated on normal and malignant cells

separately. Moreover, this proposed sensor has a high

sensitivity of 2360.12 nm/RIU, low resolution of 1.78

× 10−6, and high-quality factor (as high as 99.765)

when differentiating normal and malignant cells (P.P.

Devi and A. Panda, 2020).

Chunlian Cen et al. Here, we propose to use

critical coupling and impedance matching theory to

computationally simulate the perfect absorption of

monolayer graphene. We studied a perfect single-

band absorption of the monolayer graphene by using

the important coupling effect and impedance

matching. Errors from the data fittings remained

within 10%, leading to a high quality factor (Q-factor

= 664.2) absorption spectrum with the absorbance

~100% in the near-infrared range. The position of the

absorption spectrum can be adjusted by changing the

ratio of the air hole radii of the elliptic cylinder to

structural period. The attained S = 342.7 nm/RIU

(refractive index unit) and FOM = 199.2 (figure of

merit) could be achieved by an absorber, which shows

great potential for biosensor technology

development. We hope that our research can serve as

an interesting application for graphene photonics and

optoelectronics. Plasmon metamaterials have

recently been a hot topic due to their unique EM

(electromagnetic) control capability. As such, it is

currently the most studied material due to its optical

and physical properties. The scope of

electromagnetic metamaterials has gone from being

limited to microwave frequency range to terahertz,

infrared, and almost the entire visible light

electromagnetic spectrum as illustrated by the study

(Jiang L et., al.2020) Metamaterial absorbers, on the

other hand, have been shown to improve absorption

in solar, microwave, infrared, and optical systems.

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

740

3 EXISTING SYSTEM

To realize early detection of blood cancer, the

current paper proposes twin-core photonic crystal

fiber (TC-PCF), which involves the refractive index

(RI) of healthy and cancerous blood cells. Normal

and cancerous cells are thought to be 30 to 70 percent

liquid, according to what is known. Because the

middle air hole is longer than the other two, these

samples should be infiltrated at that spot. Fig. 1:

Schematic setup of the known TC-PCF for the normal

and malignant cells, using a two-dimensional finite

element method (FEM) to monitor the variations of

the transmitted spectrum and coupling length while

considering the refractive index (RI). As per the

transmitted spectrum shift, the proposed sensor may

show a high sensitivity (8571.43 nm/RIU). Due to the

straightforward detection strategy, the suggested TC-

PCF sensor can potentially be applied to detection of

blood cancer in a convenient and cost-effective way.

4 PROPOSED SYSTEM

This technique combines Convolutional Neural

Networks (CNN) with ResNet architecture to enhance

the analysis of blood cell images for improved

accuracy in identifying anomalies, including blood

cancers. The integration of residual learning into

ResNet ensures efficient information flow by using

shortcut connections to avoid problems such as

disappearing gradients. The classification is made

through multiple convolutional layers that help

extract features such as size, form and texture that

help distinguish between healthy blood cells and

malignant ones. Using pooling and fully connected

layers, those features are then merged to obtain a

proper classification. Honed for processing

efficiency, the system handles large, diverse datasets

and is flexible enough to perform a variety of

analytical tasks. Providing a consistent, automated

way to analyse medical images, it has the potential to

be used in multiple disease settings because of its

scalability. Automated blood smears analysis has

improved significantly with this state-of-the-art

technique, increasing diagnostic fidelity and enabling

accurate analysis, diagnosis, and prediction.

Figure 1: System Flow Diagram.

4.1 Load Data

The Load Data Module is the first step and it is used

to collect and prepare the dataset for further

operations. This module is responsible for

recovering images, which are tagged with the type of

blood cancer or whether it is non-cancerous. To

facilitate organizing and retrieving the pictures, they

are organized systematically This module ensures

that the inputs are ready in such a fashion and

structure, that the system can use it appropriately in

the subsequent workflows, which is training the data

and testing the data.

4.2 Data Preprocessing

The data preprocessing module’s objective is to clean

and transform the raw data into a form that could be

used as input in the model. This means that all the

pictures have to be resized to a standard size, their

pixels have to be transformed to the same range, as

well as performing all the other required

manipulations to get the data compatible with the

model. The objective of this module is to obtain

better quality data so that noise and inconsistencies

are discarded, and high-level structured data is clear

and organized for the best possible outcome during

the analysis and learning stages.

LOAD

DATA

PRE-

PROCESSI

MODE

EVAL

UATI

TRAINING

AND

TESTING

FEATURE

EXTRACTI

RESULT

Enhancing Blood Cancer Diognosis with Data Driven Techniques

741

4.3 Feature Extraction

This means that it is the job of the Feature Extraction

Module to find and select features of the blood sample

photos that are relevant to classification. This

module utilizes the various textures, patterns, and

shapes found in the images to recognize and classify

distinct types of blood cancer. These important

features can be automatically extracted by deep

learning technique types of computations including

convolutional neural networks (CNNs). This module

not only simplifies the data but focuses well on the

important portions of the pictures, while also helping

the model to figure, and, classify the pictures

accurately.

4.4 Training and Testing

The Module for Training and Testing incorporates the

learning process of the model. During the training

phase, the model learns to recognize patterns in the

photos based on labeled data. It adjusts its internal

parameters to minimize errors and improve

predictions. An independent set of data is used to

evaluate the model's accuracy and generalizability

after training. This ensures that the model can

accurately classify new and unseen images,

predicting how well it will perform when used on

other datasets in the future.

4.5 Model Evaluation

The fine-tuning phase of the Model Evaluation

Module determines the performance of a trained

model and tests its

accuracy. After training/testing

metrics like F1-score, accuracy, precision, recall are

used to determine

the model performance. Such

metrics assess the models ability

to classify the

benign and malignant samples and the types of blood

cancer. Make sure that the final

model is trustworthy

and gives consistent quality predictions by taking a

look at the output of the model and to find out if

improvements can be made.

5 RESULT ANALYSIS

The system's ability to detect and classify different

types of blood cancer using picture data is

demonstrated in the analysis of the project's outcome.

Performance evaluation exhibits how the ability of

the model to accurately predict benign vs malignant

samples and classify them under certain categories as

early, pre or advanced phases. Metrics of the

classification process, such as accuracy, precision,

recall, and F1-score, help ensure that the feature

extraction and model training modules have

successfully identified key patterns that help in the

human defects. The outcomes, which demonstrate

pleasingly high degrees of accuracy and consistency

in predictions, confirm the approach of analyzing

blood sample images through deep-learning

algorithms. This review illustrates how one such

initiative measured up against its goals and shows

how it might aid in diagnosis of early and precise

blood cancer type. Accuracy for the existing system

and proposed system are tabulated in table 1 and

illustrated in figure 2.

Table 1: Comparison Table.

Algorithm Accuracy

Existing system

Proposed system 85

Figure 2: Graph Diagram.

6 CONCLUSIONS

In conclusion, this study provides an approach to

employ state-of-the-art machine learning algorithms

to diagnose and classify blood cancer using image

data. By processing the data in a systematic manner,

they were able to extract relevant features, train the

model and classifier such that the blood samples were

able to identify the visual patterns in the images.

Model evaluation ensures accuracy and reliability of

the model. This method shows how successful

machine learning will be impactful in the medical

industry by providing a helpful tool to assist the

diagnosis and detection of blood-related disorders.

The method is a significant advance in medical

Accuracy

Existing system Proposed

System

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

742

analysis as it automates the identification process

and produces reliable results.

Future Work.

To increase the generalizability of the model, this

study can be further improved in the future by

including more heterogeneous datasets including

samples from other demographics and a wider

spectrum of blood cancer types. Adding in data from

additional sources, such as genetic markers or clinical

test results, could potentially improve the predictive

power of the system. Further studies could

potentially focus on enhancing feature extraction

techniques; perhaps by utilizing highly advanced

deep learning architectures ormethods such as

transfer learning to bolster model performance. Also,

we hope to reduce the computation complexity and to

tune the model for faster training steps to scale up the

system for more applications. Another potential

future direction is to automate the process itself for

various blood-related diseases. This would make the

system more useful than just cancer detection and

ultimately facilitate more accurate and efficient

diagnosis in clinical environments.

REFERENCES

A.K. Paul, M.S. Habib, N.H. Hai, S.A. Razzak, Air-core

photonic crystal fiber-based plasmonic sensor for high

refractive index sensing, Opt. Commun. 464 (2020)

125556.

Cen C, Chen X, Zhang Y, Yang H, Yi Z, Yao W, et al. A

two-band metamaterial absorber for surface plasmon

resonance at terahertz in graphene. 2020; Physica E. 10,

117:113840.

Dai, W., Chen, F., Luo, H., Xiong, Y., Wang, X., Cheng,

Y., and Gong, R. are the authors of this work. For

enhanced microwave absorption capabilities, carbonyl

iron@void@ nitrogen-doped carbon with a yolk-shell

structure is created. J. Alloy Compd. 2020, 812,

152083.

Jiang L, Xu D, Yi Z, Chen Z, Cen C, and others. Graphene

is a high quality factor, high sensitivity metamaterial

that is the perfect absorber due to critical coupling

theory and impedance matching. Nanomaterials, 2020,

10:95

M. Al Mahfuz, M.A. Hossain, E. Haque, N.H. Hai, Y.

Namihira, and Ahmed A plasmonic RI sensor with a

dual-core photonic crystal fiber that operates in the

visible to near-infrared spectrum, In IEEE Sensors

Journal (2020),

https://doi.org/10.1109/JSEN.2020.2980327

M.R.B.A. Faysal, M.R. Hasan, M.B. Sagnac

interferometer-based photonic crystal fiber salinity

sensor with high sensitivity, Results Phys. 16 (2020)

103022,

Miao, C.; Fang, R.; Mou, H.Y.; and Xiao, W. Sandwich-

like Si@TiO2@rGO composites are easily synthesized

for application as high-performance anodes in lithium

ion batteries. J. Alloy, Compd. C 2020, 818, 152884.

[Source]

P.P. Devi and A. Panda. Refractive index-based photonic

crystal biosensor for the identification of cancerous

cells, Opt. Fiber Technol. 54 (2020) 102123,

Z. Fan, Z. Guo, X. Kong, and Z. Meng, Wide- range

refractive index sensor based on photonic crystal fiber

with phase matching between metal defect mode and

core mode, Opt. Commun. (2020) 125233.

Zhang, C.F.; Xu, D.Y.; Wu, P.H.; Chen, Z.Q.; Jian, R.H.

Graphene-Based Unpatterned Narrow Dual-Band

Monolayer Perfect Absorber with Near-Infrared

Critical Coupling. Micromachines 11, 58. 8, 2020.

Enhancing Blood Cancer Diognosis with Data Driven Techniques

743