Page 74 - eBook_Proceedings of the International Conference on Digital Manufacturing V2

Proceedings of the International Conference on Digital Manufacturing –
Volume 2

The mean pixel accuracy across all foreground classes
(0.8238) highlights the model’s capability to reliably classify
pixels belonging to cervical cells of interest. However, the
significantly lower mean, Intersection over Union (IoU),
calculated at 0.0592, underscores a limitation in precise
delineation of cellular boundaries, particularly where overlapping
and ambiguous cell boundaries are prevalent. Such challenges are
typical in cytological imagery and highlight the inherent
complexity of accurate segmentation tasks. The confusion matrix
analysis in Figure 24 offers critical visualisation of classification
performance, clearly illustrating confusion patterns, especially
among Dyskeratotic, Koilocytotic, and Metaplastic classes. This
pattern of confusion identifies targeted areas for potential
improvements. Future research could include employing
specialised data augmentation techniques, refining annotation
protocols for greater boundary precision, and adopting loss
functions tailored to handle significant class imbalances, such as
focal loss.

In summary, the Swin Transformer-based semantic
segmentation model has demonstrated considerable potential for
clinical adoption in cervical cancer screening, characterised by
high recall and pixel accuracy for the most clinically critical cell
categories. However, addressing lower precision and boundary
delineation accuracy through targeted improvements and
advanced methodological strategies will be essential for
enhancing the clinical utility and diagnostic reliability of the
model.

Visualisation of Segmentation Results

To qualitatively assess the performance of the proposed model,
sample segmentation outputs are presented in Figure 27. These
visualisations depict the model's ability to localise and classify
various cervical cell types based on their morphological
characteristics. For improved interpretability, each predicted class
is assigned a distinct colour:

58