Page 59 - eBook_Proceedings of the International Conference on Digital Manufacturing V2

A NOVEL SWIN TRANSFORMER-
BASED APPROACH FOR

AUTOMATED CERVICAL CANCER
DIAGNOSIS

Imran Ali (Yuan Ze University) *, Salman Masroor (Balochistan
University of Engineering and Technology, Khuzdar) Jenson
Christopher Halim (Bina Nusantara University), Syed H. shah (Bina
Nusantara University, Yuan Ze University), Tze Chi Hsu (Yuan Ze
University), Abdullah Mengal (Balochistan University of Engineering
and Technology, Khuzdar)

ABSTRACT

Cervical cancer persists as a leading cause of female mortality
worldwide, with early detection being pivotal for effective
treatment. Current diagnostic reliance on manual Papanicolaou
(Pap) smear analysis faces challenges of subjectivity and
inefficiency. This study introduces a Swin Transformer-based
deep learning model for automated, high-precision classification
of cervical cells into normal, precancerous, and malignant
categories. The proposed architecture leverages hierarchical
windowed self-attention to capture local and global cytological
features, while Mask2Former generates interpretable
segmentation masks to elucidate model decisions for clinical
transparency. Trained on a meticulously annotated dataset of
cervical cell images (N=3,049 samples across five classes), the
model achieves outstanding recall rates of 0.96 for Parabasal and
Superficial-Intermediate cells, demonstrating exceptional
sensitivity to diagnostically critical abnormalities. Quantitative
evaluation reveals a mean pixel accuracy of 82.38%, with strong
performance in minority classes (Dyskeratotic: 77.85%,
Koilocytotic: 66.85%).

43