Page 64 - eBook_Proceedings of the International Conference on Digital Manufacturing V2

Proceedings of the International Conference on Digital Manufacturing –
Volume 2

These developments underscore a paradigm shift from hand-
engineered and CNN-based systems toward Transformer-driven
architectures, which offer superior scalability and representational
power (Riana et al., 2023; Hussain, Mahanta, Das & Talukdar,
2020). Building upon this trajectory, the Swin Transformer
emerges as a promising candidate. Its hierarchical window-based
attention mechanism enables fine-grained local feature learning
and efficient global context modelling. While the Swin
Transformer has achieved competitive results in general-purpose
vision tasks (e.g., ADE20K, COCO), its potential in cervical
cytology analysis remains largely unexplored, forming this study's
foundation.

MATERIAL & METHODOLOGY

Dataset Description

The dataset used in this study was compiled through the
conventional collection of cervical cell samples using a spatula
and brush, as illustrated in Figure 21. These cells were extracted
from the cervix and examined under a microscope to obtain high-
resolution images, which served as the foundational input for
training and evaluating the deep learning model. These images are
critical in allowing the model to learn morphological patterns
necessary for classifying cervical cells.

Figure 21: Cervical cancer diagnosis: from traditional sampling to
AI analysis

48