Histopathological Image Analysis System for Automated Breast Cancer Classification
- 주제(키워드) : Deep Learning , Histopathology images , , ResNet34 , Digital Pathology , Artificial Intelligence
- 발행기관 아주대학교
- 지도교수 Peom Park
- 발행년도 2021
- 학위수여년월 2021. 2
- 학위명 석사
- 학과 및 전공 일반대학원 시스템공학과
- 실제URI http://www.dcollection.net/handler/ajou/000000030603
- 본문언어 영어
- 저작권 아주대학교 논문은 저작권에 의해 보호받습니다.
초록/요약
From the last decade there were certain advancements seen in the field of Pathology to detect various type of diseases including cancer. Artificial Intelligence(AI) and machine learning played a main role in the development of digital slides and establishment of virtual platforms for sharing this compact data between medical professionals. However digital pathology has overtaken the traditional pathology, but this field is still in the need of a fully automated system that can handle the big data models. To get digital images; microscopic slides are scanned by high quality scanners which will give images of various formats(.png, .jpg,.tiff, .svs etc.) and sizes(kilo bytes to giga bytes). Histopathology images are usually scanned at high magnification rates and are in .tiff and .svs format of gigabyte size. Handling of such large size histopathology images is not an easy task. There is a need of a computer aided system that can analyze these images as an input directly and give results on the basis self-interpretation. Many automated cancer detection systems have been made using deep learning techniques, but size of image reduced from thousands of pixels to hundreds of pixels to fit in the deep learning model which affect the quality of an image and render output results. In this thesis I present the solution to this problem by designing classification model that can classify normal and invasive breast histopathological images . ICIAR Breast Cancer dataset has been used for training three different kind of ResNet34 models(ResNet34, ResNet 34 — With Weighted Random Sampler Model and ResNet 34 — with SMOTE Sampling Technique) using Transfer learning and achieved the highest accuracy of 96% with SMOTE Sampling Technique.
more목차
1.Introduction 1
1.1.Related Work 1
1.2.Overview 2
2. Material and Method 3
2.1. Dataset 3
2.2. System Design 4
3. Preprocessing 5
3.1.Proposed Preprocessing Model 6
4. Deep Learning Model 7
4.1.DataSplit 7
4.2.ModelTraining using Transfer Learning 7
4.2.1. ResNet34 8
4.2.2.ResNet34—With Weighted Random Sampler 8
4.2.3:ResNet34—With SMOTE Sampling Technique 9
5. Discussion and Results 9
5.1. F_score, Loss, Learning Rate Graphs and Confusion Matrix 11
5.1.1. ResNet34 11
5.1.2. ResNet34 — With Weighted Random Sampler 12
5.1.3: ResNet34 — With SMOTE Sampling Technique 13
5.2. Classification Results 14
5.3. Conclusion and Future Direction 14
References
