Page 65 - eBook_Proceedings of the International Conference on Digital Manufacturing V2

Fabrication and Characterization of a Low-Cost Piezoelectric using Rochelle
Salt for Energy Harvesting and Sensor Applications

A physician, affiliated with the Instituto de Oncologia de Ijuí
in Ijuí, Rio Grande do Sul, collaborated with the clinic to compile
a private dataset of Pap smear samples collected over several
years. The samples exhibit varied colorations and originate from
multiple microscopy techniques, including archived
examinations. A pathologist captured tissue cell images via
microscopy and annotated them using the Bethesda system.
Classification details are provided in Table 8, with representative
examples in Table 9.

Due to the high resolution and dense cellular content, the
images were cropped into smaller segments, with annotations
adjusted accordingly. The processed dataset was then uploaded to
Roboflow.com for preprocessing and augmentation.

Table 8: Private dataset
Cell Name Cancer/Normal Cells Quantity
Superficial- Normal 831
Intermediate
Parabasal 787
Metaplastic Pre-cancer - 793
Koilocytotic Abnormal 825
Dyskeratotic Cancer - Abnormal 813

Table 9: Private dataset example cells

Methodology

This study employs the Swin Transformer architecture to
accurately classify cervical cell images into three essential
diagnostic categories: normal, precancerous, and cancerous. As

49