Digimammocad: a new deep learning-based cad system for mammogram breast cancer diagnosis with mass identification
Worldwide Breast Cancer (BC) is the most severe cancer in women. There are no outward symptoms at an early stage and the survival rate decreases with the increasing stage. So, only regular screening can save a life. Mammography is the gold standard imaging modality used for regular BC screening due...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | English English English |
| Published: |
2022
|
| Subjects: | |
| Online Access: | http://eprints.uthm.edu.my/8493/1/24p%20SUSAMA%20BAGCHI.pdf http://eprints.uthm.edu.my/8493/2/SUSAMA%20BAGCHI%20COPYRIGHT%20DECLARATION.pdf http://eprints.uthm.edu.my/8493/3/SUSAMA%20BAGCHI%20WATERMARK.pdf http://eprints.uthm.edu.my/8493/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Worldwide Breast Cancer (BC) is the most severe cancer in women. There are no outward symptoms at an early stage and the survival rate decreases with the increasing stage. So, only regular screening can save a life. Mammography is the gold standard imaging modality used for regular BC screening due to its fast acquisition and cost-effectiveness. The available Computer-Aided Detection (CAD) systems based on traditional Machine Learning (ML) systems are unable to reduce the number of undetected and false-positive breast cancer cases because of their dependency on external feature extractors that provides an inferior abstraction of feature representations. Whereas Deeper Convolutional Neural Networks (DCNNs) can automatically extract features from their inputs, and hence, a remarkable change has been observed in medical image screening. A DCNN-based CAD system suffers overfitting due to the scarcity of the annotated data and the inconsistency was reported in their performance when they were validated with external datasets. So, an effective CAD system, DIGIMAMMOCAD, was developed in this study for digital mammogram screening to diagnose BC to reduce the number of false-positive and undetected cases. It was developed using a pre-trained Residual Network of 50 layers (Resnet50) and You Only Look Once (YOLO) V2 detector where only Full Field Digital Mammograms (FFDMs) were considered. The novelty of the work lies in the increased image input layer size of the Resnet50, 2 extracted feature maps for mass identification, and the use of a small dataset. The DIGIMAMMOCAD achieved the classification accuracy, sensitivity, and specificity of 98.33 %, 0.97, and 1, respectively, along with the Average Precision (AP) of 0.91 for mass identification with the INbreast dataset. It outperformed the state-of-the-art DCNN-based systems proposed by other authors. It also achieved high performance with external datasets of different image quality after minimum adaptation in the system and reduced the false-positive and undetected cases remarkably. So, the DIGIMAMMOCAD can have a significant contribution to clinical use, and can also serve the medical fraternity as well as patients in a better way. |
|---|