Scribd
Upload
15 views

Breast Cancer Detection Using Deep Learning

Uploaded by

lumini wickremesinghe
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
15 views

Breast Cancer Detection Using Deep Learning

Uploaded by

lumini wickremesinghe
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 20

Breast Cancer Detection using

Deep Learning

Author
Goron Paula-Melisa
CANCER
Disease of the century
A medical condition that can affect any
part of the body
May be caused by an abnormal growth
of cells

Reasons
DETECTION
It helps with early cancer detection by
identifying anomalies in medical images
Detection improves patients' health and
saves lives
SOURCE: A COMPETITION, BENCHMARK, CODE, AND DATA FOR USING ARTIFICIAL
INTELLIGENCE TO DETECT LESIONS IN DIGITAL BREAST TOMOSYNTHESI

DATASET
22032 FILES
training 19148
validation 1163
test 1721
220 images with bounding boxes for malignant or benign lesions.

RESULTS
Team 7: The model's ability to detect correctly is 0.786, meaning it successfully identified 78.6% of individuals

Related
with breast cancer. With 95% confidence, there is some level of uncertainty associated with the estimate, and
the model's detection capability can range from 0.720 to 0.852."

work Prezentările sunt instrumente de comunicare ce pot fi utilizate ca demonstrații.


CNN

YOLO

DATASET
Solution
PREPROCESSING

EXPERIMENTS

EVALUATION
CONVOLUTIONAL LAYER

It uses filters to scan the input image and extract important features.

POOLING
Reduces the number of network parameters
Max-pooling selects only the most significant values from a region.
Preservs important features.
Creates a more compact representation of the image.

FULLY CONNECTED LAYER

CNN These layers take all the previous features and process them to generate final results.
Often followed by a final activation layer (e.g., softmax).
YOLO

YOLO (You Only Look Once) - object detection


It utilizes a neural network to predict bounding
boxes and object classes.
Sizes: YOLOv5s, YOLOv5m, YOLOv5l, YOLOv5x -
Differ in complexity and performance.
It utilizes a CNN architecture
DBT - BREAST CANCER SCREENING -
DIGITAL BREAST TOMOSYNTHESIS
(BREAST-CANCER-SCREENING-DBT)
Dataset
FILE-PATHS.CSV
pacient id RADIOLOGY MANUAL
study id ANNOTATIONS.CSV
view name of the pathologies
path of the images image benign
malign
LABELS.CSV pacient id
normal
labels
benign RADIOLOGY_HAND_DRAWN
malign _SEGMENTATIONS_V2.CSV
normal bounding box : X, Y, Width, Height
BOXES.CSV CDD-CESM - CATEGORIZED DIGITAL
(220 RECORDS)
DATABASE FOR LOW ENERGY AND
bounding box Slice
SUBTRACTED
X
CONTRAST ENHANCED SPECTRAL
Y
MAMMOGRAPHY IMAGES
Width
Height
800 images
Source: Cancer Imaging Archive (TCIA)
DATASET Preprocessing
train - 80%
validation - 10% DBT AND CDD-CESM
test - 10% (x, y)- (x, y) - the coordinates of the top-left corner of
the bounding box,
width, height of the bounding box.
YOLO
YOLO
Each image - label (x, y) - the normalized coordinates of the center point
txt file: of the bounding box

X normalized width, height - the normalized width and height of the


bounding box
Y normalized
width normalized
nx = (x+Width/2)/width−of −image
height normalized
Weights
Experiments

WEIGHTS DROPOUT
pre-trained parameters of the YOLO model (transfer learning) used to improve the performance of neural network models
obtained by training the model on a large training dataset (e.g., and prevent overfitting
COCO - Common Objects in Context). it randomly deactivates a number of neurons (or connections
vanilla - no pre-trained parameters are used between neurons) with a predefined probability

OPTIMIZER
automatically adjusts the weights to minimize a loss (or cost)
function
BOXLOSS
OBJLOSS
Evaluation
CLSLOSS
measures how well the model
locates objects in an image.
measures the accuracy of the
model's object detection within a
METRICS/RECALL
specific area (grid cell). measures how many classes were
measures the error in classifying predicted correctly out of the total
objects. number of existing classes

METRICS/PRECISION
measures how many classes were
METRICS/MAP_0.5
predicted correctly out of the total
a metric that considers the
number of predicted classes
performance of Bbox localization
(through Intersection over Union)
and the model's classification ability
by combining precision and recall
Evaluation
5th Experiment

6th Experiment
1st Experiment
Labels Predictions
Evaluation
2nd Experiment
Labels Predictions
Evaluation
4th Experiment
Labels Predictions
Evaluation
5th Experiment
Labels Predictions
Evaluation
6th Experiment
Labels Predictions
Evaluation
Strengths:
good predictions
portability
easy detection on other datasets
SWOT Weaknesses:
Analysis few images
Opportunities:
usable in the medical field

Threats:
similar models
using SGD as optimiser is better than
using Adam
a larger dataset significantly
Conclusions improves the model's performance
using pre-trained weights provides
better results than vanilla
expanding/Diversifying the dataset

adding background images


Future
work increasing the image size for training

choosing a different optimizer algorithm


(e.g., Momentum)
Buda, Mateusz et al., Breast Cancer Screening - Digital Breast Tomosynthesis
(Breast-Cancer-Screening-DBT), The Cancer Imaging Archive, 2020, DOI:
10.7937/E4WT-CD02.
Ferreira, João Otávio N., Orange Trees Detection with YOLOv5 in UAV Imagery, URL:
https://joaootavionf007.medium.com/orange-trees-detection-with-yolov5-in-
uav-imagery-22ec29db922e.
Konz, N., & Buda, M. (2023). A Competition, Benchmark, Code, and Data for Using
Artificial Intelligence to Detect Lesions in Digital Breast Tomosynthesis. JAMA
Network Open, 6(2), e230524. doi:10.1001/jamanetworkopen.2023.0524.
References Li, Fei-Fei s, i Karpathy, Andrej, “Convolutional Neural Networks for Visual
Recognition”, în Stanford University (2015).
Nielsen, Michael A., Neural Networks and Deep Learning, Determination Press, 2015.
Raschka, Sebastian, “A Gentle Introduction to Backpropagation”, în Blog post
(2018).
The Cancer Imaging Archive, The Cancer Imaging Archive, National Cancer
Institute, URL: https://www.cancerimagingarchive.net/.
Ultralytics LLC, YOLOv5, https://github.com/ultralytics/yolov5.
https://www.upgrad.com/blog/basic-cnn-architecture/

You might also like