Page 77 - 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

DISCUSSION

While the Swin Transformer-based semantic segmentation model
demonstrated strong potential for cervical cell classification,
several critical limitations emerged during evaluation. A major
challenge was the class imbalance within the dataset—classes
such as Dyskeratotic and Koilocytotic were significantly
underrepresented compared to more prevalent categories, like
Parabasal and Superficial-Intermediate.

This imbalance led to lower precision scores and reduced the
model’s ability to consistently distinguish between
morphologically similar cell types, resulting in increased false
positives. Another observed limitation was the difficulty in
precise cell boundary delineation, largely due to the complex and
overlapping nature of cervical cytology images. This issue was
reflected in the relatively low Intersection over Union (IoU)
scores, particularly for minority classes, indicating challenges in
capturing detailed boundaries during segmentation. To mitigate
these challenges, several strategies are proposed.

Targeted data augmentation, including techniques such as
Generative Adversarial Networks (GANs) or class-balanced
oversampling, can help increase the representation of minority
classes and improve model generalisation. Additionally, improved
annotation protocols that emphasise clearer cell boundary
labelling may enhance segmentation accuracy. From a training
standpoint, adopting specialised loss functions such as focal loss
or class-weighted cross-entropy can help the model to focus more
effectively in underrepresented categories, reducing false
positives and improving overall precision.

CONCLUSION AND FUTURE WORK.

In conclusion, the Swin Transformer-based semantic
segmentation model demonstrates robust potential for accurately
classifying cervical cell images, emphasising clinically significant
cancerous and precancerous categories.

61