Application of Multi-View Swin Transformer on Mammogram Images for Breast Cancer Diagnosis Through XAI-Based BI-RADS and Breast Density Classification

Oktafrian, Muhamad Vicky (2026) Application of Multi-View Swin Transformer on Mammogram Images for Breast Cancer Diagnosis Through XAI-Based BI-RADS and Breast Density Classification. Undergraduate thesis, UPN Veteran Jawa Timur.

	Text (Cover) 22081010028,-cover.pdf Download (575kB)
	Text (Bab 1) 22081010028.-bab1.pdf Download (165kB)
	Text (Bab 2) 22081010028.-bab2.pdf Restricted to Repository staff only until 30 July 2029. Download (806kB)
	Text (Bab 3) 22081010028.-bab3.pdf Restricted to Repository staff only until 30 July 2029. Download (828kB)
	Text (Bab 4) 22081010028.-bab4.pdf Restricted to Repository staff only until 30 July 2029. Download (680kB)
	Text (Bab 5) 22081010028.-bab5.pdf Download (154kB)
	Text (Daftar Pustaka) 22081010028.-daftarpustaka.pdf Download (197kB)

Abstract

Breast cancer is among the cancers with the highest incidence and mortality rates worldwide, making early detection through mammography and BI-RADS and breast density classification critically important. However, manual interpretation of mammograms by radiologists is prone to subjectivity, while most existing deep learning models still rely on single-view images and lack integration with XAI. This study aims to apply a Multi-View Swin Transformer architecture to classify BI-RADS and breast density using ipsilateral pairs of Cranio Caudal and Medio Lateral Oblique images, and to evaluate the interpretability of Grad-CAM visualizations. The dataset used is VinDr-Mammo (20,000 images from 5,000 patients), which underwent preprocessing including intensity scaling, CLAHE, Min-Max normalization, and resizing to 224×224, then was split 80:20 using stratified sampling. Features from both views were fused through an Omni-Attention mechanism, with five class-imbalance handling schemes (baseline, class weights, oversampling, undersampling, and synthetic data augmentation) each tested under four optimizer (Adam and AdamW) and batch size (8 and 16) configurations, resulting in 40 experimental scenarios evaluated using accuracy and AUC. The results show that for BI-RADS classification, the oversampling pipeline achieved the best performance with an AUC of 0.7834, while for breast density classification, the baseline model without imbalance handling performed best with an AUC of 0.9462. Grad-CAM visualizations demonstrate that the model is able to focus on clinically relevant image regions. The best-performing model was subsequently deployed into a Streamlit-based web application equipped with out-of-distribution detection using Mahalanobis Distance.

Item Type:

Thesis (Undergraduate)

Contributors:

Contribution	Contributors	NIDN/NIDK	Email
Thesis advisor	Puspaningrum, Eva Yulia	NIP198907052021212002	evapuspaningrum.if@upnjatim.ac.id
Thesis advisor	Muttaqin, Faisal	NIP198512312021211009	faisalmuttaqin.if@upnjatim.ac.id

Subjects:

Q Science > QA Mathematics > QA75 Electronic computers. Computer science

Divisions:

Faculty of Computer Science > Departemen of Informatics

Depositing User:

Muhamad Vicky Oktafrian

Date Deposited:

30 Jul 2026 04:22

Last Modified:

30 Jul 2026 04:22

URI:

https://repository.upnjatim.ac.id/id/eprint/58310